KR20110123770A - Image identifier extraction device - Google Patents

Abstract

This image identifier extracting apparatus is provided with image identifier generating means and encoding means. The image identifier generating means includes one or more partial region pairs in which both of a pair of shapes of paired subregions and a relative positional relationship of two paired partial regions are different from at least one other partial region pair. According to the plurality of partial region pairs in the image, the region feature amount is extracted from each partial region of the image, and an image identifier used for identification of the image is generated based on the extracted region feature amount of each partial region. do. The encoding means encodes the picture identifier.

Description

Image identifier extraction device {IMAGE IDENTIFIER EXTRACTION DEVICE}

The present invention relates to a system for extracting an image identifier which is a feature amount (for determining identity) for identifying an image.

An image identifier is an image feature amount (for determining identity) for identifying an image. By comparing the image identifier extracted from one image and the image identifier extracted from another image, it is possible to calculate an identity measure (typically referred to as similarity or distance) in which two images show the same degree from the comparison result. In addition, it is possible to determine whether two images are the same by comparing the calculated identity measure with a threshold value. Here, "the two images are the same" is not limited to the case where the two images are the same at the level of the image signal (pixel value of the pixels constituting the image). Changing the aspect ratio, adjusting the color tone of the image, processing various filters on the image (sharpening, smoothing, etc.), local processing on the image (caption superposition, cutout, etc.), recapturing the image, etc. It also includes the case where one image is a duplicated image of the other image by the various alteration processes of. By using the image identifier, for example, the duplication of an image or a moving image which is a set of images can be detected, so that the image identifier can be applied to an illegal copy detection system or the like of an image or a moving image.

An example of an image identifier is described in patent document 1. As shown in FIG. FIG. 18 is a diagram illustrating a method for extracting an image identifier described in Patent Literature 1. FIG. This picture identifier is a feature vector of a plurality of dimensions (16 dimensions in FIG. 18). Average luminance values are respectively calculated from 32 rectangular regions 244 at predetermined positions in the image 240 (in FIG. 18, 16 rectangular regions are drawn), and the paired rectangular regions (paired rectangular regions in FIG. 18) are respectively calculated. The difference in the average luminance value is calculated on the dotted line 248) to obtain a 16-dimensional difference vector (250). A composite vector is generated for the difference vector 250 by vector transform, and the 16-dimensional quantization index vector obtained by quantizing each dimension of the composite vector is used as an image identifier.

Prior art literature

Patent Literature

Patent Document 1: Japanese Patent Publication No. Hei 8-5500471

[0005] An image identifier composed of feature vectors of a plurality of dimensions has a smaller (or less redundancy) information amount of a feature vector as the correlation between the dimensions becomes smaller, so that an identification capability that is enough to identify different images is obtained. Increases. On the contrary, when the correlation between the dimension of the feature vector is large, the amount of information the feature vector has is small (large redundancy), so that the identifying ability is low. Here, the correlation between dimensions is the degree of similarity of occurrence of the feature quantities of the dimension, and mathematically, for example, the correlation coefficient between the probability variables and the mutual information amount when the occurrence of the feature quantities of each dimension is used as the random variable. It is a value calculated as. For this reason, it is preferable that the image identifier composed of the feature vectors of a plurality of dimensions be designed so that the correlation between the dimensions is small.

Image signals (pixel values of pixels constituting an image) have a correlation between local regions of an image. In general, the closer the distance between local regions is, the greater the correlation. In particular, for example, an image in which a certain image pattern appears repeatedly (especially when the image pattern appears repeatedly at regular intervals) (for example, an image of a window of a building arranged in a grid pattern, etc., see FIG. 19A) Or the image (refer to FIG. 19B) which consists of a specific texture, etc., the correlation between local areas of an image becomes large.

[First Problem]

An image identifier composed of a feature vector composed of feature quantities extracted from a plurality of local regions of an image, as described in Patent Literature 1, has a feature amount in each dimension for an image having a large correlation between local regions of the image. Since the shape of the local area | region which extracts this is the same (the rectangular area of the same shape in the example of patent document 1), the correlation between the dimension of the feature amount to be extracted becomes large. For this reason, there is a first problem that the identification capability of the picture identifier (feature vector) is lowered. Note that the same shape here means the same, including the size and angle of the area (tilt or orientation).

For example, about an image in which a certain image pattern repeatedly appears (see FIG. 19A), an image composed of a specific texture (see FIG. 19B), and the like, as described in Patent Literature 1 The image identifier has low identification capability.

[Second Problem]

The second problem of the image identifier described in Patent Document 1 is that since the shape (including size and angle) of the area of each dimension for calculating the feature amount (feature vector) is the same rectangle, the length of the side of the rectangle There is a blind spot on the frequency such that a frequency component having a period equal to or one of its integers cannot be detected. The reason for this is that if an average is taken within a region of a signal component of this particular frequency, it becomes 0 regardless of the magnitude of the signal component, and the signal of the frequency component cannot be detected at all. More specifically, assuming that a frequency having a period equal to the length of a rectangular side is f ₀ , the component of frequency nf _{0 (} n = 1, 2,3, ...) cannot be detected. For this reason, the average value of the pixel values becomes the same as the direct current component with respect to the direct current component and the image in which the signal is concentrated on this frequency component, and the difference in value between the regions is eliminated. As a result, the value of all the feature amounts extracted as the difference between the average pixel values between the regions becomes 0 and cannot be identified (the identification capability is significantly reduced). In reality, it is difficult to detect not only the components of the frequency nf _{0 (} n = 1, 2,3, ...) but also the constant frequency region in the vicinity thereof, even if the signal components are not concentrated on the specific frequency. Signal component can not be used, and the identifying ability is lowered. In order to alleviate this problem, it is conceivable to increase the value of the frequency f ₀ to reduce the signal power falling into the difficult-to-detect frequency band. However, increasing the value of the frequency f ₀ means reducing the size of the region, which leads to a deterioration in the robustness of the characteristic amount (the degree to which the characteristic amount does not change with respect to various modification processing and noise). For example, when the area becomes small, the value of the feature amount changes significantly even for a small position shift, and the robustness of the feature amount is lowered. Thus, when using the same rectangular area, it is very difficult to secure robustness while increasing the discrimination ability.

[Objective of the invention]

SUMMARY OF THE INVENTION An object of the present invention can solve the problem of a deterioration in the identification ability of the image having a large correlation between local regions of an image or an image identifier extracted from an image in which a signal is concentrated at a specific frequency. It is to provide an image identifier extraction device.

[0011] Such an image identifier extracting apparatus according to one aspect of the present invention includes at least one portion in which both a combination of shapes of two partial regions to be paired with and a relative positional relationship between the two partial regions to be paired are different from each other. According to the plurality of partial region pairs in the image, which include one or more partial region pairs different from the region pairs, region feature amounts are extracted from each partial region of the image, and based on the region feature amounts for each extracted partial region. Image identifier generating means for generating an image identifier for use in identification of the image; And encoding means for encoding the image identifier.

Since the present invention is configured as described above, it is possible to increase the identification ability, which is a degree capable of identifying different pictures of the picture identifier. In particular, this effect is remarkable for an image having a large correlation between local regions of the image.

In addition, according to the present invention, there is an effect that the identification capability does not decrease even for an image in which signals are concentrated at a specific frequency.

1 is a block diagram of a first embodiment of the present invention.
2 is a diagram illustrating an example of a pair of extraction areas for each dimension in which the extraction information for each dimension is shown.
3 is a block diagram showing an example of a comparison means in the first embodiment of the present invention.
4 is a block diagram showing another example of the comparison means in the first embodiment of the present invention.
Fig. 5 is a flowchart showing the flow of processing in the first embodiment of the present invention.
6 is a block diagram of an essential part of a second embodiment of the present invention.
7 is a flowchart showing the flow of processing in the second embodiment of the present invention.
8 is a block diagram of a third embodiment of the present invention.
9 is a diagram illustrating an example of a method for calculating region feature quantities for each dimension.
Fig. 10 is a flowchart showing the flow of processing in the third embodiment of the present invention.
Fig. 11 is a block diagram showing a fourth embodiment of the present invention.
12 is a table showing an example of the comparison and quantization method for each dimension.
It is a flowchart which shows the flow of a process of 4th Embodiment of this invention.
Fig. 14A is a table showing dimensional extraction area information used in the sixth and seventh embodiments of the present invention.
FIG. 14B is a table showing dimension-specific extraction region information used in the sixth and seventh embodiments of the present invention. FIG.
Fig. 14C is a table showing dimensional extraction region information used in the sixth and seventh embodiments of the present invention.
FIG. 14D is a table showing dimension-specific extraction region information used in the sixth and seventh embodiments of the present invention. FIG.
FIG. 14E is a table showing dimension-specific extraction region information used in the sixth and seventh embodiments of the present invention. FIG.
FIG. 14F is a table showing dimension-specific extraction region information used in the sixth and seventh embodiments of the present invention. FIG.
Fig. 14G is a table showing dimensional extraction area information used in the sixth and seventh embodiments of the present invention.
FIG. 14H is a table showing dimension-specific extraction region information used in the sixth and seventh embodiments of the present invention. FIG.
FIG. 14I is a table showing dimension-specific extraction region information used in the sixth and seventh embodiments of the present invention. FIG.
FIG. 14J is a table showing dimension-specific extraction region information used in the sixth and seventh embodiments of the present invention. FIG.
It is a table | surface which shows the area characteristic amount calculation method information for each dimension used by 6th Embodiment of this invention.
It is a table | surface which shows the area characteristic amount calculation method information for each dimension used by 6th Embodiment of this invention.
It is a table | surface which shows the dimension characteristic area calculation method information for each dimension used by 6th Embodiment of this invention.
Fig. 15D is a table showing the area characteristic amount calculation method information for each dimension used in the sixth embodiment of the present invention.
It is a table | surface which shows the dimension characteristic area calculation method information for each dimension used by 6th Embodiment of this invention.
It is a table | surface which shows the information of the method of calculating the area | region characteristic quantity for each dimension used by 7th Embodiment of this invention.
FIG. 16B is a table showing dimension information for each region feature amount calculation method used in the seventh embodiment of the present invention. FIG.
Fig. 16C is a table showing dimension information of each region feature amount calculation method used in the seventh embodiment of the present invention.
It is a table | surface which shows the area characteristic amount calculation method information for each dimension used by 7th Embodiment of this invention.
Fig. 16E is a table showing the area characteristic amount calculation method information for each dimension used in the seventh embodiment of the present invention.
FIG. 17A is a table showing dimensional comparison and quantization method information used in the sixth and seventh embodiments of the present invention. FIG.
FIG. 17B is a table showing dimensional comparison and quantization method information used in the sixth and seventh embodiments of the present invention. FIG.
FIG. 17C is a table showing dimensional comparison and quantization method information used in the sixth and seventh embodiments of the present invention. FIG.
FIG. 17D is a table showing dimension comparison and quantization method information used in the sixth and seventh embodiments of the present invention. FIG.
FIG. 17E is a table showing dimensional comparison and quantization method information used in the sixth and seventh embodiments of the present invention. FIG.
It is a figure which shows the extraction method of the image identifier described in patent document 1. FIG.
Fig. 19 is a diagram illustrating an example of an image in which the correlation between local regions becomes large.
20 is a block diagram showing a fifth embodiment of the present invention.
Fig. 21 is a block diagram of a combining means for performing a combination between quantization index vectors.
Fig. 22 is a flowchart showing an example of a process of combining means for performing combination between quantization index vectors.
Fig. 23 is a flowchart showing another example of the processing of the combining means for performing the combination between the quantization index vectors.
Fig. 24 is a flowchart showing another example of the processing of the combining means for performing the combination between the quantization index vectors.
Fig. 25 is a block diagram showing another configuration of the combining means for performing the combination between the quantization index vectors.
Fig. 26 is a flowchart showing an example of a process of combining means for performing combination between quantization index vectors.
Fig. 27 is a flowchart of another processing example of the combining means for combining the quantization index vectors.
Fig. 28 is a diagram showing an example of an index given to 1024 blocks that can divide an image into 32 vertical directions and 32 horizontal directions.
It is a figure which shows the area | region which belongs to one type among the area | region corresponding to each dimension of 8th Embodiment of this invention.
Fig. 29B is a table showing an area belonging to one type of areas corresponding to each dimension of the eighth embodiment of the present invention.
Fig. 29C is a table showing an area belonging to one type of areas corresponding to each dimension of the eighth embodiment of the present invention.
Fig. 29D is a table showing areas belonging to one type of areas corresponding to each dimension of the eighth embodiment of the present invention.
Fig. 29E is a table showing a region belonging to one type of regions corresponding to each dimension of the eighth embodiment of the present invention.
Fig. 29F is a table showing a region belonging to one type of regions corresponding to each dimension of the eighth embodiment of the present invention.
Fig. 29G is a table showing an area belonging to one type of areas corresponding to each dimension of the eighth embodiment of the present invention.
FIG. 30 is a table showing a relationship between an area type of each dimension, the number of dimensions, and an index corresponding to a threshold value.
It is a figure which shows an example of the 1st, 2nd extraction area | region of the dimension of area | region type a.
It is a figure which shows an example of the 1st, 2nd extraction area | region of the dimension of area | region type b.
Fig. 31C is a diagram illustrating an example of the first and second extraction regions in the dimension of the region type c.
It is a figure which shows an example of the 1st, 2nd extraction area | region of the dimension of area | region type d.
It is a figure which shows an example of the 1st, 2nd extraction area | region of the dimension of area | region type e.
It is a figure which shows an example of the 1st, 2nd extraction area | region of the dimension of area | region type f.
It is a figure which shows an example of the 1st, 2nd extraction area | region of the dimension of area | region type g.

[First Embodiment]

[Configuration of First Embodiment]

Next, 1st Embodiment of this invention is described in detail with reference to drawings.

Referring to FIG. 1, the image identifier extracting apparatus according to the first embodiment of the present invention includes an image identifier as a feature vector (more specifically, a quantization index vector) having a plurality of dimensions with respect to an input image. As a system that outputs. The image identifier extracting apparatus includes a dimension determining means 1, an extracting region obtaining means 2, an area feature variable calculating means 3, and a comparing means 4.

The dimension determining means 1 next determines the dimension of the feature vector to be extracted and supplies it to the extraction region acquisition means 2. The dimension determination means 1 sequentially supplies the dimension of the feature vector to extract, and the component after the extraction area | region acquisition means 2 extracts the feature quantity corresponding to the supplied dimension. For example, when the feature vector is configured in N dimensions, the dimension determination means 1 may sequentially supply the extraction region acquisition means 2 from the first dimension to the Nth dimension. Finally, once all the dimensions of the feature vector are supplied, they can be supplied in any order of dimensions. Plural dimensions may be supplied in parallel.

The extraction region acquisition means 2 is supplied with dimension-specific extraction region information as an input separately from the dimension supplied from the dimension determination means 1.

The dimension-specific extraction region information is information indicating a predetermined pair of the first extraction region and the second extraction region, which are associated for each dimension of the feature vector. The first and second extraction regions have the following characteristics as essential conditions.

[Required Conditions of First and Second Extraction Areas]

The prerequisites of the first and second extraction regions are that the relative positions of the extraction region pairs differ among the dimensions, and that the combination of the shape of the extraction region pairs among the dimensions differs.

For each of the dimensions indicated by the dimension-specific extraction information, an example of a pair of extraction regions that satisfy the above essential condition is shown in FIG. 2. It differs from the extraction area of the image identifier shown in FIG. 18, and the knitting of the shape of the pair of extraction areas is different among each dimension. Here, different shapes may refer to the same shape at different angles (for example, the second extraction region of the first dimension and the first extraction region of the seventh dimension) of FIG. 2, or similar shapes of different sizes (eg, A second extraction region of the first dimension of FIG. 2 and a second extraction region of the ninth dimension). In addition, it is a minimum condition that at least one dimension in which the pair of extraction areas has a combination of different shapes is included in the entire dimension of the feature vector. It is preferable that more dimensions have different shapes (combination) of pairs of extraction regions in the feature vector. This is because the more the dimensions of different shapes (combination) of the pairs of extraction regions are different from each other, the smaller the correlation between the dimensions than those of the feature vector, and the higher the discrimination ability. For example, the shapes of the pairs of extraction regions may differ from one another among all dimensions of the feature vector.

The first extraction region and the second extraction region in the dimension do not have to be the same shape as in the ninth dimension of FIG. 2, and the shapes may be different as in the other dimensions of FIG. 2. When the shape of the first extraction region and the second extraction region in each dimension is different, the correlation between the feature amount extracted from the first extraction region and the second extraction region is small, and the identification capability is increased. This is therefore desirable. In addition, since the possibility of the first extraction region and the second extraction region becoming the frequency blind spot at the same frequency is lowered at the same time, the identification capability is increased.

The shape of each extraction region is arbitrary. For example, any complex shape, such as the second extraction region of the sixth dimension of FIG. 2, may also be acceptable. If the extraction region consists of a plurality of pixels of the image, line segments or curves may also be allowed, for example, as in the seventh or tenth dimension of FIG. 2. Further, for example, a plurality of small regions in which the extraction regions are not continuous, such as the first extraction region in the eighth dimension, the first and second extraction regions in the eleventh dimension, and the first extraction region in the twelfth dimension It may be composed of regions. As such, when the feature vector includes an extraction region of a complicated shape, the correlation between the dimensions of the feature quantities extracted therefrom can be lowered, so that the identifying ability can be increased.

For example, as in the fifth dimension of FIG. 2, a part of the first extraction region and the second extraction region may overlap each other. In addition, any one of the extraction region pairs may be contained in the other. As such, by allowing redundancy in the pair of extraction regions, more patterns (relative positions, distances) can be taken in the extraction region pairs, so that patterns that can reduce the correlation between dimensions can be increased, thereby improving identification ability. The probability of making it higher is increased.

In addition, different from the extraction area of the image identifier shown in FIG. 18, as in each of the dimensions shown in FIG. 2, the extraction areas may partially overlap each other. As shown in the extraction area of the image identifier shown in Fig. 18, if the extraction area is exclusively taken between the dimensions, the possible patterns of the extraction area pairs are limited. As shown in Fig. 2, by allowing the overlapping of the extraction regions between the dimensions, the pattern that can reduce the correlation between the dimensions can be increased, thereby increasing the possibility of making the identification ability higher. However, if there are too many overlapping portions of the extraction area between the dimensions, the correlation between the dimensions becomes large and the identification ability becomes low. Therefore, this is not desirable.

In addition, in the case of integrating the extraction areas of all dimensions, the extraction area is taken so that the area where the feature amount is not extracted in the image is reduced (that is, almost the full screen of the image can be covered). It is preferable. As shown in Fig. 18, when a large number of regions in which the feature amount is not extracted is included in the image, much information included in the image signal (pixel values of pixels constituting the image) is not used, and the identification capability is not increased. . In the case of incorporating the extraction areas of all dimensions, it is included in the image signal by taking the extraction area so that the area in which the feature amount is not extracted in the image becomes smaller (that is, almost the full screen of the image can be covered). As more information can be reflected in the feature amount, the identification ability can be enhanced. In addition, when the extraction areas of all dimensions are integrated, it is preferable that the extraction areas are not deflected and uniformly acquired from the entire image. However, if there is a high probability that local processing such as subtitle superimposition is performed on any particular region, it is preferable to obtain the extraction region avoiding the region. In addition, since the peripheral region of the edge of the image often does not include the feature portion of the image, it is preferable to obtain the extraction region avoiding the peripheral region.

In addition, it is preferable that the size, relative position (distance, direction) of the extraction region follow a constant distribution (for example, a uniform distribution), because the relative position (distance, direction) is according to a uniform distribution, This is because the extraction area is not deflected with respect to the distance or direction, so that the extraction area does not concentrate at a specific distance or direction, and thus more diversity can be achieved. In addition, the closer the relative position is, the larger the correlation between the regions is, so that the larger the difference in shape is, the better the relative position is.

The extraction region information for each dimension may be any form of information as long as the first extraction region and the second extraction region for each dimension are uniquely identifiable information. Since the extraction area must always be the same area for any size or aspect ratio image, the dimension-specific extraction area information can be obtained in the same extraction area for any size or aspect ratio image. It should be information of. For example, the dimension-specific extraction region information describes the position and shape of the extraction region with respect to an image having a predetermined size and aspect ratio (for example, an image having a width of 320 pixels × a width of 240 pixels). You may. In this case, for an image input at any arbitrary size and aspect ratio, the image is first resized to have a predetermined size and aspect ratio, and then according to the position and shape of the extraction region described in the extraction region information for each dimension. The extraction region may be specified. Conversely, the extraction region may be specified by converting the position and shape of the extraction region described in the dimensional extraction region information corresponding to the image of the arbitrary size and aspect ratio of the input image.

Information representing each extraction region included in the dimension-specific extraction region information is, for an image of a predetermined size and aspect ratio (for example, an image of 320 pixels wide × 240 pixels wide), It may be information describing a set of coordinate values of all the pixels constituting the extraction area. The information indicating each extraction region included in the dimension-specific extraction region information may be information that describes, as parameters, the position and shape of the extraction region with respect to an image of a predetermined size and aspect ratio. For example, if the shape of the extraction area is a rectangle, the information may describe the coordinate values of the four corners of the rectangle. In addition, when the shape of the extraction area is a circle, the information may describe the coordinate value and the radius value of the center of the circle.

In addition, by using the seed of the pseudo random number as the extraction area information for each dimension, starting from the species in the extraction area acquisition means 2 to generate a pseudo random number, extraction of different shapes according to the random number You can also choose to create an area (for example, the four corners of a rectangle are determined by random numbers). Specifically, for example, the extraction region for each dimension can be obtained in the following order.

 (1) A seed of pseudo random number is supplied as the extraction area information for each dimension.

 (2) The dimension is set to n = 1.

 (3) A pseudo random number is generated to determine four corners of the rectangle of the first extraction region of the dimension n.

 (4) A pseudo random number is generated to determine four corners of the rectangle of the second extraction region of the dimension n.

 (5) The dimension is set to n = n + 1, and the procedure returns to (3).

Since the extraction region is determined based on the random number, the generated extraction region has a different shape for each dimension. In addition, since the same random number is generated every time (for any image) when the pseudo random number is the same, the same extraction area is reproduced for the different images.

The extraction region acquisition means 2 is information indicating a first extraction region and a second extraction region corresponding to the dimension supplied from the dimension determination means 1 from the dimension-specific extraction region information supplied as an input. Is obtained and supplied to the extraction region representative value calculation means (3).

In addition to the input from the extraction area acquisition means 2 (information indicating the first extraction area and the second extraction area), the area feature variable calculating means 3 serves as an input of the image identifier to be extracted. The image is supplied. The area feature variable calculating means 3 has a first area feature variable calculating means 31 and a second area feature variable calculating means 32. The area feature variable calculating means 3 uses the first area feature variable calculating means 31 to extract the first extraction region supplied from the extraction area acquisition means 2 for each dimension from the image supplied as the input. Based on the information to be shown, the feature amount of the first extraction region is calculated as the first feature feature and supplied to the comparison means 4. Moreover, the area feature variable calculating means 3 extracts the 2nd feature supplied from the extraction area acquisition means 2 for every dimension from the image supplied as an input using the 2nd area feature variable calculating means 32. As shown in FIG. Based on the information indicating the area, the feature amount of the second extraction area is calculated as the second area feature amount and supplied to the comparison means 4.

Further, in order to specify each extraction area for the input image based on the information indicating the first extraction area and the second extraction area, the area feature variable calculating means 3, if necessary, The image is resized so as to have a predetermined size and aspect ratio of the dimensional extraction area information.

The area feature variable calculating means 3 calculates the area feature variable of each extraction region by using pixel values of the pixel group included in each extraction region. Here, the pixel value is a value of a signal that each pixel of the image has, and is a scalar amount or a vector amount. For example, when the image is a luminance image, the pixel value is a luminance value (scalar amount), and when the image is a color image, the pixel value is a vector amount indicating a color component. For example, when a color image is an RGB image, a pixel value is a three-dimensional vector amount of R component, G component, and B component. When the color image is a YCbCr image, the pixel value is a three-dimensional vector amount of the Y component, the Cb component, and the Cr component.

In order to calculate the area feature amount of the extraction area, the method of calculating the extraction area (the first extraction area and the second extraction area) of the dimension is constant (the same calculation method is used for any input image). ), Any method may be used.

The area feature amount to be calculated may be a scalar amount or a vector amount. For example, when the pixel value is a scalar amount such as a luminance value, the area feature amount may be calculated as an average value, a median value, a maximum number of bins, a maximum value, a minimum value, or the like (all are scalar amounts). For example, the pixel values included in the extraction region may be sorted, and the pixel values at positions of a predetermined ratio from the upper or lower portion of the distribution (sorted permutation) may be calculated as region feature amounts (this is also a scalar amount). ). More specifically, the case where P% of the percentage is a predetermined ratio (for example P = 25%) will be described. The pixel values (luminance values) of the N pixels in the extraction area are sorted in ascending order, and the set of pixel values (luminance values) sorted in ascending order is Y (i) = {Y (0), Y ( 1), Y (2),... , Y (N-1)}. Here, the pixel value at the position of P% from the lower part of the permutation sorted in ascending order becomes, for example, Y (floor (N × P / 100)), and this value is obtained as the area feature amount of the extraction region. do. In addition, floor () is a function that performs truncation after the decimal point. Here, the region feature amount calculated by applying this formula (Y (floor (N × P / 100))) to the luminance value of the pixel included in the extraction region is referred to as "percentile luminance value feature amount". .

In addition, when the pixel value is a vector amount such as a color component, first, they may be converted into a scalar amount by any method, and then the area feature amount may be calculated by the method described above. For example, when the pixel value is a three-dimensional vector amount of RGB components, first, the area values may be calculated by the method described above after converting them into luminance values that are scalar amounts. In addition, when the pixel value is a vector amount, it is also possible to use, for example, an average vector of pixel values included in the extraction area as the area feature amount.

For example, you may perform arbitrary calculations (differential calculation, filtering operation), such as edge detection and template matching, with respect to an extraction area | region, and may use the calculation result as area | region feature amount. For example, it may be a two-dimensional vector amount indicating the direction of the edge (direction of the gradient), or a scalar amount indicating similarity with any template.

In addition, a histogram representing the color distribution included in the extraction region, the directional distribution of the edges, and the intensity distribution of the edges may be obtained as the area feature amounts (all are vector quantities).

In addition, various feature quantities defined in ISO / IEC 15938-3 may be used, which are dominant color, color layout, scalable color, color structure. Edge Histogram, Homogeneous Texture, Texture Browsing, Region Shape, Contour Shape, Shape 3D, Parametric Motion Motion, and Motion Activity.

The comparison means 4 compares the first area feature amount supplied from the area feature variable calculating means 3 with the second area feature amount for each dimension, and quantizes the result of comparison to obtain the quantization index. Outputs Since the comparison means 4 outputs a quantization index for each dimension, a quantization index vector consisting of quantization indices of a plurality of dimensions is finally output.

The comparison means 4 may use any method for comparing and quantizing the first region feature amount with the second region feature amount. The number of quantization indexes for each dimension is also arbitrary.

The comparison means 4, for example, when the area feature amount is a scalar amount (for example, an average value of luminance values), and compares the magnitude and the quantization index is +1 when the first area feature amount is large. In the case of, the quantization index may be set to −1, and the comparison result may be quantized to a quantization index of two values of +1 and −1. Here, if the first region feature amount of the dimension n is Vn1 and the second region feature amount is Vn2, the quantization index Qn of the dimension n can be calculated by the following equation.

[Equation 1]

Qn = + 1 (for Vn1> Vn2)

   -1 (for Vn1≤Vn2)

Here, FIG. 3 shows a more detailed configuration diagram of the comparison means 4 in the case where the comparison means 4 performs the comparison and quantization based on Equation 1 above.

Referring to FIG. 3, the comparison means 4 includes a magnitude comparison means 41 and a quantization means 42.

When the first area feature amount and the second area feature amount are supplied, the magnitude comparison means 41 compares the value of the first area feature amount with the value of the second area feature amount, and compares the result. Is supplied to the quantization means 42. This compares the magnitude of Vn1 and Vn2 with the magnitude comparison means 41, and supplies the quantization means 42 with the information which shows whether a comparison result is Vn1> Vn2 or Vn1 <= Vn2.

The quantization means 42 quantizes according to Equation 1 based on the result of the magnitude comparison supplied from the magnitude comparison means 41, and outputs a quantization index. Therefore, when the information indicating that the comparison result is Vn1> Vn2 is supplied, the quantization means 42 causes the quantization index to be +1, while the information indicating that the comparison result is Vn1? Vn2 is supplied. The quantization index is output such that the quantization index is -1.

In addition, note that the comparison and quantization method based on the equation 1 will be referred to as the comparison and quantization method A.

In addition, the comparison means 4, for example, when the area feature amount is a scalar amount (for example, an average value of the luminance values), when the absolute value of the difference value is equal to or less than a predetermined threshold value, It is determined that there is no difference between the amount and the second area feature amount, and a quantization index 0 indicating that there is no difference is set, and in other cases, the comparison means 4 compares the magnitude with the first area. If the feature amount is larger, the quantization index +1 is set. Otherwise, the quantization index -1 is set, and the quantization index is quantized in such a manner that the quantization index is any one of three values of +1, 0, -1. It can also be done. Here, assuming that the first region feature amount of the dimension n is Vn1, the second region feature amount is Vn2, and the predetermined threshold is th, the quantization index Qn of the dimension n can be calculated by the following equation.

[0052] [Equation 2]

Qn = + 1 (when | Vn1-Vn2 |> th and Vn1> Vn2)

    0 (for | Vn1-Vn2 | ≤th)

   -1 (when | Vn1-Vn2 |> th and Vn1≤Vn2)

4 shows a more detailed configuration diagram of the comparison means 4 in the case where the comparison means 4 performs the comparison and quantization based on the equation (2).

Referring to FIG. 4, the comparing means 4 includes a difference value calculating means 43 and a quantization means 44. The quantization means 44 is previously supplied with a threshold value, which is information (quantization boundary information) that is predetermined, indicating the boundary of quantization.

The difference value calculating means 43, if the first region feature amount and the second region feature amount are supplied, the difference between the value of the first region feature amount and the value of the second region feature amount. Value is calculated and the calculated difference value is supplied to the quantization means 44. This means that the difference value calculating means 43 calculates Vn1-Vn2 and supplies the resultant value to the quantization means 44.

The quantization means 44 is based on equation 2 based on a threshold value which is information (quantization boundary information) indicating the boundary of the difference value supplied from the difference value calculation means 43 and the predetermined quantization boundary supplied as an input. Quantization is performed accordingly, and a quantization index is output. This is based on the value of Vn1-Vn2 supplied from the difference calculating means 41 and the threshold value th supplied as an input, and the quantization means 42 quantizes in the case of | Vn1-Vn2 |> th and Vn1-Vn2> 0. If the index is +1, | Vn1-Vn2 |> th, and Vn1-Vn2≤0, the quantization index is -1, and if | Vn1-Vn2 | ≤th, the quantization index is output in such a manner that the quantization index is set to 0. do.

In addition, a comparison and quantization method based on this equation 2 will be referred to as a comparison and quantization method B hereinafter.

Although quantization is performed here at three values based on the difference, it is possible to perform quantization at a more (level) quantization index, depending on the magnitude of the difference. Also in this case, the comparison means 4 has the structure shown in FIG. 4, and in the quantization means 44, a plurality of threshold values as information (quantization boundary information) indicating a predetermined boundary of quantization of each level are used as input. Supplied. In addition, a comparison and quantization method of quantizing at four or more levels based on the difference value and the plurality of threshold values supplied as an input will be referred to as a comparison and quantization method C below.

As such, the difference between the first area feature amount and the second area feature amount is small (below a predetermined threshold value), indicating that there is no difference when it is determined that the difference does not exist. By introducing a quantization index, compared with the method of Equation 1, the feature amount (quantization index) of the dimension of the pair of the extraction regions where the difference between the area feature amounts is smaller is more stable, i.e., for various modification processing and noise. It can be hardened. Therefore, an image identifier (quantization index vector) that is stable even for a flat image (e.g., a blue sky image) that has a small overall difference in characteristics between local regions and that is robust against various alteration processes and noise is obtained. You can print

Further, the comparison means 4 may, for example, when the area feature amount is a vector amount, first convert the vector amount to a scalar amount by any method, and then perform quantization by the method described above (the comparison And the quantization method will be referred to as comparison and quantization method D below. For example, you may calculate the difference vector which is a difference with the vector of a 2nd extraction area | region from the vector of a 1st extraction area | region, vector-quantizes a difference vector, and may compute a quantization index. In this case, for example, a predetermined representative vector (such as a center vector) is supplied for each quantization index so that it is classified as a quantization index having the largest similarity (smallest distance) between the representative vector and the difference vector ( This comparison and quantization method will be referred to as comparison and quantization method E). In addition, similarly to the quantization of the scalar amount according to Equation 2 described above, when the norm of the difference vector is equal to or less than a predetermined threshold, there is no difference between the first region feature amount and the second region feature amount. By judging, a quantization index indicating no difference may be introduced as quantization index 0 indicating no difference.

In addition, when matching the quantization index vector output in the present invention (when comparing the quantization index vector extracted from the image and the quantization index vector extracted from another image to determine whether or not their image is the same), quantization The number of dimensions (similarity) in which the indexes match or the number of dimensions (hamming distance) in which the quantization indexes are inconsistent are calculated as the identity measure, and the calculated identity measure is compared with a threshold to determine the identity of the image. Moreover, in the comparison means 4, when the quantization index is calculated based on Formula 2, the identity measure (similarity) can be calculated as follows. First, the quantization index vectors of two images are compared with each other between corresponding dimensions to calculate the number of dimensions in which both quantization indexes are not 0 (this value is set to A). Next, in the dimension where both quantization indices are not 0, the number of dimensions with which the quantization indices match is calculated (set this value to B). The similarity is calculated as B / A. Here, in the case of A = 0 (ie, when both quantization indices become zero in all dimensions), the similarity is set to a predetermined numerical value (for example, 0.5).

[Operation of the First Embodiment]

Next, with reference to the flowchart of FIG. 5, operation | movement of the image identifier extraction apparatus in 1st Embodiment is demonstrated. In the flowchart of FIG. 5, the dimension (number) of the feature vector is represented by “n”, and the dimension includes a total N dimension from 1 to N. FIG.

First, the dimension determination means 1 determines dimension 1 as the first dimension for extracting the feature vector (n = 1), and supplies it to the extraction region acquisition means 2 (step A1).

Next, the extraction region acquisition means 2 obtains information indicating the first extraction region and the second extraction region of the dimension n from the dimensional extraction region information supplied as an input, and the region feature amount. It supplies to the calculating means 3 (step A2).

[0065] Next, the area feature variable calculating means 3 calculates the first area feature variable and the second area feature variable of the dimension n from the image supplied as an input, and supplies it to the comparison means 4. (Step A3).

Next, the comparison means 4 compares the first region feature amount and the second region feature amount of the dimension n, quantizes the comparison result, and outputs a quantization index (step A4).

[0067] Next, it is determined whether or not the output of the quantization index has ended for all the dimensions (ie, it is determined whether n <N is true or false) (step A5). If the output of the quantization index has ended for all dimensions (that is, n <N is true), the process ends. If the output of the quantization index has not ended for all dimensions (i.e., n <N is true), the process proceeds to step A6. In step A6, the dimension determination means 1 determines the next dimension for extracting the feature vector (n = n + 1), and supplies it to the extraction region acquisition means 2. The process then proceeds to step A2 again.

In addition, although extraction processing is performed in order from dimension 1 to dimension N here, it is not limited to this order and arbitrary orders can be taken. Moreover, not only this process sequence but extracting process with respect to several dimension can also be performed in parallel.

[Effects of the First Embodiment]

Next, the effect of 1st Embodiment of this invention is demonstrated.

[0070] The first effect is that it is possible to increase the identifying ability of the picture identifier composed of the feature vectors of the plurality of dimensions, which is the extent to which different pictures can be identified. In particular, this effect is remarkable for an image having a large correlation between local regions of the image.

The reason is that the correlation between the dimensions can be reduced as the shapes of the regions from which the feature amounts are extracted from the dimensions are different (variable shapes of the regions).

The second effect is that the identification capability will not be degraded even for the image in which the signal concentrates on a specific frequency.

The reason is that as the shape of the region extracting the feature quantities between the dimensions is different (various shapes of the region), even for an image in which the signal concentrates at a certain frequency, all (many) extraction regions at the same time This is because there is no difference between the feature quantities of the pairs (dimensions) of H, which makes it difficult to occur when the discrimination ability is lowered.

[Second Embodiment]

[Configuration of Second Embodiment]

Next, 2nd Embodiment of this invention is described in detail with reference to drawings.

In a second embodiment of the present invention, the comparison means 4 of the first embodiment shown in FIG. 1 is replaced with the comparison means 4A shown in detail in FIG. It differs from embodiment. Since components other than the comparison means 4A are the same as in the first embodiment, the description of such components is omitted here.

Referring to FIG. 6, the comparison means 4A includes a difference value calculating means 43, a quantization boundary determining means 45, and a quantization means 44.

The difference value calculating means 43 calculates a difference value between the first area feature amount and the second area feature amount supplied from the area feature variable calculating means 3 for each dimension, and determines the quantization boundary determination means. To 45 and the quantization means 44.

[0078] The difference value is obtained by subtracting a second area feature amount (or vice versa) from the first area feature amount, for example, when the area feature amount is a scalar amount (for example, an average value of luminance values). The amount of scalar In addition, when the area feature amount is a vector amount, for example, after converting each vector to a scalar amount by any method, a difference value of the scalar amount may be obtained. In addition, when the area feature amount is a vector amount, the difference vector between the first area feature amount and the second area feature amount may be referred to as a difference value (vector amount).

If the difference values for all the dimensions of the feature vector supplied from the difference value calculation means 43 are supplied to the quantization boundary determination means 45, the quantization boundary determination means 45 is based on the distribution of the difference values of all the dimensions. The boundary of the quantization is determined, and the determined quantization boundary related information is supplied to the quantization means 44. Here, the distribution of the difference values of all dimensions means the frequency (probability) of occurrence with respect to the difference value (or difference vector).

Determining the boundary of quantization means determining parameters for assigning the quantization index to the quantization index without exception and exclusively when quantizing the difference value. When the difference value is a scalar amount, for example, a range (that is, a threshold value) for each quantization index (quantization level) is determined, and the range (threshold value) is supplied to the quantization means 43 as information on the quantization boundary. do. When the difference value is a vector amount, for example, a parameter for performing vector quantization, that is, a representative vector (center vector, etc.) of each quantization index is determined, and the quantization means 44 is used as information on the quantization boundary. )

[0081] The quantization boundary determining means 45, based on the distribution of the difference values of all dimensions, when the difference value is a scalar amount and quantizes M values (M = 2, 3, ..., etc.) You may determine the range (threshold) of quantization so that the ratio with respect to the whole dimension of an index is equal.

For example, using the constant α as a modification of Equation 1, the quantization index is +1 when Vn1 + α> Vn2, and the quantization index is −1 when Vn1 + α ≦ Vn (M = 2). ), The central point of the distribution of the difference value (the point where the integral of the left and right distributions are equal) may be determined as the threshold value α of the quantization so that the ratio of the quantization index +1 and the quantization index -1 is equal. . Similarly, when the difference value is a vector amount, when quantization of M values is performed, each quantization index is equalized based on the distribution of the difference vector of all dimensions so that the ratio of the respective quantization indexes to all dimensions is equal. You may determine the area | region of the vector space to be allocated, or determine the representative vector (center vector etc.) of each quantization index at the time of vector quantization. As such, by equalizing the ratio of the quantization indices to all dimensions (ie, by removing the deflection of the quantization indices), entropy can be increased, so that the identification capability can be improved.

In addition, the quantization boundary determination means 45 determines the boundary of the quantization such that the ratio of the quantization index to the entire dimension is equalized, and the quantization means 44 performs quantization based on the determined boundary, and The quantization method will be referred to as the comparison and quantization method F.

For example, the quantization boundary determination means 45 has a difference in the case where the difference value is a scalar amount and the quantization of three values by Equation 2 (the quantization index is +1, 0, -1). The threshold value th (used to set the quantization index to 0 if it is below this threshold value) used to quantize to the quantization index 0 indicating no is determined based on the distribution of the difference values of all dimensions, and the determined threshold value th is determined. You may supply to 44 (in the comparison means 4 shown in FIG. 4 of 1st Embodiment, this threshold th is set previously). For example, the absolute value of the difference value of all the dimensions is calculated, the absolute value of the calculated difference value is sorted, and from the upper or lower level, the point of a predetermined ratio (this predetermined ratio is supplied as an input, for example) is obtained. The threshold th may be set (this comparison and quantization method is called comparison and quantization method G). In addition, the threshold th may be determined not by a predetermined ratio but by a method in which the ratio of the quantization indexes of +1, 0, -1 is equally close (this comparison and quantization method is referred to as the comparison and quantization method H). To call). The comparison and quantization method H corresponds to the specific example of the comparison and quantization method F performed according to Formula 2.

[0085] A more specific method of comparison and quantization method G is described by way of example, where the predetermined percentage is P% (eg P = 25%). The absolute values of the differences of all dimensions (number of dimensions = N) are sorted in ascending order, and the set of absolute values of the difference values sorted in ascending order is D (i) = {D (0), D (1), D (2 ),… , D (N-1)}. Here, the value at the position of P% from the lower part of the permutation sorted in ascending order becomes, for example, D (floor (N × P / 100)), and the threshold value th = D (floor (N × P / 100)). Note that floor () is a function that performs truncation after the decimal point.

The method in this embodiment can be contrasted with the case where the comparison means 4 take the configuration shown in FIG. 4, as in the first embodiment. In the configuration shown in FIG. 4 of the first embodiment, the predetermined threshold value th is supplied as an input, whereas in the above-described method of the second embodiment, the quantization boundary determining means 45 makes a difference of all dimensions. Based on the distribution of the values, the threshold value th is adaptively calculated for the image. Thus, while the threshold th is fixed in the first embodiment, the threshold th is adaptively calculated for the image in the second embodiment. Since the threshold th is adaptively calculated for the image, the value of the dimension of the feature vector is biased (high probability of occurrence of a specific quantization index) compared to the case where the threshold th is fixed. Since it can be suppressed (especially with respect to an image with few ups and downs), an identification ability can be made high. For example, when the fixed threshold value th is used as in the first embodiment, the quantized index becomes zero in the majority of dimensions (or all dimensions) of the feature vector. However, using the adaptive threshold value th of the second embodiment, the threshold value th is automatically adjusted to a small value for low-flip images, so that the quantization index becomes zero in the majority dimension of the feature vector. Things don't happen.

The quantization means 44 performs quantization based on the difference value of each dimension supplied from the difference value calculation means 43 and the information of the quantization boundary supplied from the quantization boundary determination means 45, Output the quantization index.

In addition, since the quantization means 44 becomes meaningless when quantization ignoring the quantization boundary information output from the quantization boundary determination means 45, the quantization boundary determination means 45 determines the quantization boundary. Note that it is necessary to follow the quantization method assumed.

[Operation of Second Embodiment]

Next, with reference to the flowchart of FIG. 7, the operation | movement of the image identifier extraction apparatus in 2nd Embodiment is demonstrated. In the flowchart of Fig. 7, the dimension (number) of the feature vector is represented by &quot; n &quot;, and the total N dimension from 1 to N exists.

First, the dimension determination means 1 determines dimension 1 as the first dimension for extracting the feature vector (n = 1) and supplies it to the extraction region acquisition means 2 (step B1).

Next, the extraction region acquisition means 2 acquires information indicating the first extraction region and the second extraction region of the dimension n from the region-specific extraction region information supplied as an input, and the region feature amount. It supplies to the representative value calculating means 3 (step B2).

Next, the area feature variable representative value calculating means 3 calculates the first area feature amount and the second area feature amount of the dimension n from the image supplied as an input, and calculates the difference value calculation means ( 43) (step B3).

Next, the difference value calculation means 43 calculates a difference value between the first area feature amount and the second area feature amount of the dimension n, and the quantization boundary determination means 45 and the quantization means 44 ) (Step B4).

[0094] Next, it is determined whether or not the processing up to the calculation of the difference values for all the dimensions is finished (that is, whether n <N is true or false) (step B5). When the process up to the difference value calculation for all the dimensions is finished (that is, n <N is false), the process proceeds to step B7. If the processing up to the difference value calculation for all the dimensions is not finished (that is, n <N is true), the process proceeds to step B6. In step B6, the dimension determination means 1 determines the next dimension for extracting the feature vector (n = n + 1) and supplies it to the extraction region acquisition means 2. Then, the flow proceeds to step B2 again.

In addition, although the extraction process is performed in order from dimension 1 to the dimension N here, it is noted that the order is not limited to this and may be taken arbitrarily.

[0096] Next, when the difference values of all the dimensions of the feature vector supplied from the difference value calculating means 43 are supplied, the quantization boundary determination means 45 determines the boundary of the quantization based on the distribution of the difference values of all the dimensions. It determines and supplies the information of the determined quantization boundary to the quantization means 44 (step B7).

[0097] Next, in step B8, dimension 1 is set (n = 1) as the first dimension to quantize (calculate the quantization index).

[0098] Next, the quantization means 44 performs quantization based on the difference value of the dimension n and the quantization boundary supplied from the quantization boundary determination means 45, and outputs a quantization index (step B9).

[0099] Next, it is determined whether the output of the quantization indexes for all the dimensions is finished (that is, whether n <N is true or false) (step B10). The process ends when the output of the quantization indexes for all dimensions is finished (that is, when n <N is false). If the output of the quantization indexes for all the dimensions is not finished (that is, n <N is true), the process proceeds to step B11. In step B11, the following dimension is set as the dimension of the feature vector to be quantized (n = n + 1). Then, the flow proceeds to step B9 again.

In addition, although the quantization process is performed in order from the dimension 1 to the dimension N here, it is noted that the order is not limited to this and may be taken arbitrarily.

[Effects of the Second Embodiment]

In the second embodiment, compared with the first embodiment in which the boundary of quantization is fixed, the point in which the boundary of quantization is adaptively (dynamically) calculated for the image is different. As in the first embodiment, when the boundary of quantization is fixed, for a specific image (e.g., a flat image with low relief), the value of the dimension of the feature vector is biased to a specific quantization index (specific quantization). A situation where a high probability of the appearance of an index occurs occurs (entropy is lowered), which causes a problem that the discriminating ability of these images is degraded. On the other hand, as in the second embodiment, the boundary of the quantization is adaptively (dynamically) calculated with respect to the image, so that for any image, the value of the dimension of the feature vector is biased to a specific quantization index (specific quantization). It is possible to suppress work having a high probability of appearing in the index, so that the identification ability can be improved.

[Third Embodiment]

[Configuration of Third Embodiment]

Next, 3rd Embodiment of this invention is described in detail with reference to drawings.

Referring to FIG. 8, in the third embodiment of the present invention, the area feature variable calculating means 5 is added to the configuration of the first embodiment shown in FIG. (3) This is different in that it is replaced by the area feature variable calculating means 3A having the first and second area feature variable calculating means 31A and 32A. In addition, about the other structure, since it is the same as the structure of 1st Embodiment, description is abbreviate | omitted here. In addition, although it demonstrates as knitting with 1st Embodiment here, knitting with 2nd Embodiment is also acceptable.

The region feature variable calculation method acquisition means 5 is supplied with the dimension and dimension region feature variable calculation method information from the dimension determination means 1.

[0105] The dimension feature amount calculation method information for each dimension is information indicating a method for calculating the region feature amount in the dimension, which is associated for each dimension of the feature vector, and it is essential that the method for calculating the region feature amount is different between the dimensions. to be. Note that the difference in the method for calculating the area feature amounts also includes the case where different parameters (threshold values, etc.) are applied to the same order.

Here, the area feature variable calculation method includes, for example, various methods described in the description of the area feature variable calculating means 3 of the first embodiment, and the parameters thereof.

[0107] Further, the area feature variable calculation method for each dimension indicated by the dimension feature feature calculation method information for each dimension has a minimum condition in which at least one pair of different dimensions is included in the area feature variable calculation method in all dimensions of the feature vector. Pay attention to The more the dimension of the region feature variable calculation method is different from each other, the more preferable. The more the dimension of the region feature variable calculation method is different from each other, the smaller the correlation between many dimensions in the feature vector and the higher the identification capability. For example, the area feature variable calculation methods may be different among all the dimensions of the feature vector.

[0108] In addition, the information indicating the method for calculating the area feature variable for each dimension may take any form as long as the method for calculating the area feature variable is uniquely specified.

9 shows an example of an area feature variable calculation method for each dimension. As shown in Fig. 9, the method for calculating region feature amounts differs between dimensions. In addition, as shown in the example shown in FIG. 9, the feature amounts of the scalar amount and the vector amount may be mixed. Dimension is the vector quantity).

[0110] The area feature variable calculating method obtaining means 5 is information indicating the area feature variable calculating method associated with the dimension supplied from the dimension determining means 1 from the dimension feature variable calculating method information for each dimension supplied as an input. Is obtained and supplied to the area feature variable calculating means 3A.

[0111] The area feature variable calculating means 3A is an area feature based on information indicating the first extraction area and the second extraction area supplied from the extraction area acquisition means for each dimension from the image supplied as an input. According to the information indicating the area feature variable calculation method supplied from the quantity calculation method acquiring means 5, the feature amount of the first extraction region and the feature amount of the second extraction region are respectively determined by the first region feature amount and the first value. It calculates as 2 area | region feature quantities, and supplies them to the comparison means 4.

[0112] In the area feature variable calculating means 3A, the dimension of the information indicating the supplied extraction area and the dimension of the information indicating the area feature variable calculating method need to be synchronized.

[Operation of Third Embodiment]

Next, with reference to the flowchart of FIG. 10, operation | movement of the image identifier extraction apparatus in 3rd Embodiment is demonstrated. In the flowchart of FIG. 10, the dimension (number) of the feature vector is denoted by "n", and there are total N dimensions from 1 to N. FIG.

[0114] First, the dimension determination means 1 determines dimension 1 as the first dimension for extracting the feature vector (n = 1), and the extraction area acquisition means 2 and the area feature variable calculation method acquisition means ( 5) (step C1). Next, the extraction region acquisition means 2 acquires information indicating the first extraction region and the second extraction region of the dimension n from the dimension-specific extraction region information supplied as the input, and calculates the region feature variable calculating means ( 3) Supply to A (step C2).

[0115] Next, the area feature variable calculating method obtaining means 5 obtains information indicating the area feature variable calculating method corresponding to the dimension n from the dimension-specific area feature variable calculating method information supplied as the input and the area. It supplies to the feature-quantity calculating means 3A (step C3).

[0116] Next, the area feature variable calculating means 3A calculates the first area feature variable and the second area feature variable of the dimension n from the image supplied as an input, and supplies it to the comparison means 4. (Step C4). Next, the comparison means 4 compares the first region feature amount in the dimension n with the second region feature amount, quantizes the result of the comparison, and outputs a quantization index (step C5). Next, it is determined whether or not the output of the quantization index has ended for all the dimensions (step C6). If the output of the quantization index has ended for all dimensions, the process ends. If the output of the quantization index has not finished for all the dimensions, the process proceeds to step C7. In step C7, the dimension determination means 1 determines the next dimension for extracting the feature vector (n = n + 1) and supplies it to the extraction region acquisition means 2 and the region feature variable calculation method acquisition means 5. do. The process then goes back to step C2.

In addition, although extraction processing is performed in order from dimension 1 to dimension N here, it is not limited to this, Arbitrary order may be taken. Moreover, not only this process sequence but extracting process with respect to several dimension can also be performed in parallel. In addition, the order of step C2 and step C3 may be reversed.

[Effect of Third Embodiment]

In addition to the effects of the first embodiment, the third embodiment has an effect of further increasing the discrimination ability, which is a degree capable of identifying different images.

This is because the correlation between the dimensions can be made smaller because the method for calculating the region feature amount between the dimensions is different (various region feature amount calculation methods are used).

[Fourth Embodiment]

[Configuration of Fourth Embodiment]

Next, the 4th Embodiment of this invention is described in detail with reference to drawings.

Referring to FIG. 11, the configuration of the fourth embodiment of the present invention is compared with the configuration of the first embodiment shown in FIG. 1, and a comparison method acquisition means 6 is added, and the comparison means 4 is compared. It is different in that it is replaced by the means 4B. In addition, about the other structure, since it is the same as the structure of 1st Embodiment, description is abbreviate | omitted here. In addition, although knitting with 1st Embodiment is demonstrated here, knitting with 2nd Embodiment and 3rd Embodiment is also acceptable.

[0122] The comparison method acquisition means 6 is supplied with the dimension and dimension-by-dimensional comparison method information from the dimension determination means 1.

[0123] Dimensional comparison and quantization method information is information indicating a method of comparing quantization by comparing region feature quantities in a dimension, which are associated with each dimension of a feature vector, and it is essential that the comparison and quantization methods are different between dimensions. Condition. It is also noted here that different comparison and quantization methods also include applying different parameters (threshold value, quantization index number, etc.) for the same order.

[0124] Here, the comparison and quantization method is a method of various comparison and quantization described in the description of the comparison means (4) of the first embodiment, the parameters (threshold value, number of quantization indexes, etc.) according thereto, and the second The various comparison and quantization methods described in the description of the comparison means 4A of the embodiment of the present invention, and the parameters (threshold value, quantization index number, etc.) according thereto are included.

[0125] In addition, the comparison and quantization method for each dimension indicated by the dimension comparison and quantization method information has the lowest condition that at least one pair of different dimensions must be included in the comparison and quantization methods in all dimensions of the feature vector. The more dimensions with which the comparison and quantization methods are different from each other, the more preferable, since the more dimensions with which the comparison and quantization methods are different from each other, the smaller the correlation between many dimensions in the feature vector, the higher the identification capability. Note, for example, that the comparison and quantization methods may differ from one another between all dimensions of the feature vector.

[0126] The information indicating the comparison and quantization methods for each dimension may take any form as long as the method for comparing and quantizing the area feature amounts is uniquely specified.

12 shows an example of the comparison and quantization method for each dimension. As shown in FIG. 12, the comparison and quantization methods differ between dimensions. In addition, different parameters (threshold th) may be set in the same comparison and quantization methods as in the third, fifth, and twelfth dimensions. In addition, the example of the comparison and quantization method for every dimension shown in FIG. 12 is connected with the example of the area characteristic amount calculation method for every dimension shown in FIG. Thus, the method for comparing and quantizing the scalar amount is shown as an example for the region feature amount of the scalar amount, and the method for comparing and quantizing the vector amount is shown as an example for the region feature amount of the vector amount.

[0128] The comparison method acquisition means 6 acquires information indicating a comparison and quantization method corresponding to the dimension supplied from the dimension determination means 1 from the dimension comparison and quantization method information supplied as an input, and compares the comparison means ( 4B).

[0129] The comparison means 4B compares and quantizes the first region feature amount and the second region feature amount supplied from the region feature variable calculating means 3 for each dimension from the comparison method acquisition means 6. According to the information indicating the method, the comparison and quantization are performed, and a quantization index is output. The comparison means 4B may have the structure which includes both the comparison means 4 of 1st Embodiment and the comparison means 4B of 2nd Embodiment as needed by a comparison and quantization method. have.

In the comparison means 4B, the dimension of the supplied region feature amount and the dimension of the information indicating the comparison and quantization method need to be synchronized.

[0131] [Operation of Fourth Embodiment]

Next, with reference to the flowchart of FIG. 13, operation | movement of the image identifier extraction apparatus in 4th Embodiment is demonstrated. In the flowchart of FIG. 13, the dimension ((number)) of a feature vector is represented by "n", and there exists a total N dimension from 1 to N. In FIG.

[0132] First, the dimension determining means 1 determines dimension 1 as the first dimension for extracting the feature vector (n = 1), and supplies it to the extraction region acquisition means 2 and the comparison method acquisition means 6. (Step D1). Next, the extraction region acquisition means 2 acquires information indicating the first extraction region and the second extraction region of the dimension n from the dimension-specific extraction region information supplied as the input, and calculates the region feature variable calculating means ( 3) (step D2).

[0133] Next, the comparison method acquisition means 6 obtains information indicating the comparison and quantization method corresponding to the dimension n from the dimension comparison and quantization method information supplied as an input and supplies it to the comparison means 4B. (Step D3).

[0134] Next, the area feature variable calculating means 3 calculates the first area feature variable and the second area feature variable of the dimension n from the image supplied as an input, and supplies it to the comparison means 4B. (Step D4). Next, the comparison means 4B compares the first region feature amount in the dimension n with the second region feature amount, quantizes the result of the comparison, and outputs a quantization index (step D5). Next, it is determined whether or not the output of the quantization index has ended for all the dimensions (step D6). If the output of the quantization index has ended for all dimensions, the process ends. If the output of the quantization index has not finished for all the dimensions, the process proceeds to step D7. In step D7, the dimension determination means 1 determines the next dimension for extracting the feature vector (n = n + 1), and supplies it to the extraction region acquisition means 2 and the comparison method acquisition means 6. The process then proceeds to step D2 again.

In addition, although extraction processing is performed in order from dimension 1 to dimension N here, it is not limited to this order and arbitrary orders may be taken. Moreover, not only this process order but extracting process with respect to several dimension may be performed in parallel. In addition, the order of step D2 and step D3 may be reversed, and step D3 may be performed immediately before step D5.

[Effect of Fourth Embodiment]

In addition to the effects of the first embodiment, the fourth embodiment can further increase the identification capability, which is a degree capable of identifying different images.

The reason is that, as the comparison and quantization methods are different between dimensions (various comparison and quantization methods are used), the correlation between the dimensions can be made smaller.

 [Fifth Embodiment]

[Configuration of Fifth Embodiment]

Next, a fifth embodiment of the present invention will be described in detail with reference to the drawings.

Referring to FIG. 20, the fifth embodiment of the present invention differs in that the encoding means 7 is added to the configuration of the first embodiment shown in FIG. In addition, about the other structure, since it is the same as the structure of 1st Embodiment, description is abbreviate | omitted here. In addition, although demonstrated as a knitting with 1st Embodiment here, knitting with 2nd Embodiment, 3rd Embodiment, or 4th Embodiment is also acceptable.

[0140] The encoding means 7 encodes the quantization index vector supplied from the comparison means 4 in a uniquely decodable format so that the data amount is small, and outputs the encoded quantization index vector.

[0141] The encoding means 7 may encode the data amount small, for example, by collectively encoding a plurality of dimensions instead of encoding each dimension of the quantization index vector independently.

[0142] Here, when the quantization index is calculated based on the equation 2, the encoding means 7 refers to a method for efficiently encoding. When the quantization index is calculated based on Equation 2, the quantization index of each dimension receives any one of three values of (+1, 0, -1). In the case of encoding each dimension independently, two bits (= 4 states) are required for each dimension. Here, considering that the 5 dimensions are collectively encoded (the 5 dimensions may be, for example, an arbitrary combination including 5 consecutive dimensions), the number of states becomes a power of 3 = 243, and 1 Byte = 8 bits (= 256 states) can be represented (within 256 states). In that case, the average number of bits required for one dimension is 8/5 = 1.6 bits, so that the data amount can be made smaller than in the case of encoding each dimension independently (0.4 bits per dimension can be reduced). . For example, when the total number of quantization index vectors is 300 dimensions, when each dimension is encoded independently, 2 bits x 300 = 600 bits = 75 bytes. On the other hand, if the data is collectively encoded for every five dimensions, 1.6 bits x 300 = 480 bits = 60 bytes, and 15 bytes can be reduced.

[0143] A specific example of encoding the state of three values of (+1, 0, -1) calculated in accordance with Expression 2 for every five dimensions is shown below. The combination of each 5-dimensional set may be any combination, but there is, for example, a coding method for each successive 5-dimensional set. That is, the fifth dimension can be collectively encoded from the first dimension, the tenth dimension can be collectively encoded from the sixth dimension, and the fifteenth dimension can be collectively encoded from the eleventh dimension. May be a combination of two dimensions). Here, if the values of the five-dimensional quantization indexes to be coded together are set to Q _n , Q _{n +1} , Q _{n +2} , Q _{n +3} , and Q _{n + 4} (each takes one of +1, 0, -1). For example, the encoded value Z can be calculated according to the following equation.

[0144] [Equation 3]

Z = (3 ⁴ × Q _n +1)} + {3 ³ × Q _{n +1} +1)} + {3 ² × Q _{n +2} +1)} + {3 ¹ × Q _{n +3} +1)} + {3 ⁰ × Q _{n +4} +1)}

Since the encoded value Z takes a value from 0 to 242 (243 state), the value of the quantization index is encoded as one byte (8 bits) of data. In addition, a method of mapping the values of the five-dimensional quantization index to be collectively coded from Q _n , Q _{n +1} , Q _{n +2} , Q _{n +3} , Q _{n +4} from 0 to 242 (243 states) is given by the following equation. 3] is not limited only. For different combinations of five-dimensional quantization indices, any method can be used as long as it is a method that maps to a different value (value of 243 states). Based on the equation given in [Equation 3], the mapping (value after encoding) may be calculated, encoded or generated and stored in the mapping table in advance, and the mapping (value after encoding) is obtained while referring to the stored correspondence table. May be encoded.

[0146] The method of efficiently encoding when the quantization index described in paragraph 0145 to paragraph 0145 is calculated based on Equation 2 is not limited to the case where the quantization index is calculated based on Equation 2. Is equally applicable if is a quantization index vector of a state of three values. That is, when a quantization index vector consists of quantization indexes of three values, five dimensions can be collectively encoded as 1 byte = 8 bits. Since five-dimensional quantization indexes having three-value states can have 243 different combinations of five-dimensional quantization indices, each byte is mapped to a value of 0 to 242 (243 states) so that 1 byte = 8 bits. Can be encoded. In addition, this mapping may calculate and encode a mapping (value after encoding) based on a given expression as shown in [Equation 3], or generate and store a mapping table of mapping in advance, and refer to the stored mapping table while mapping (value after encoding). ) May be obtained and encoded.

[0147] As described above, instead of encoding each dimension of the quantization index vector independently, the plurality of dimensions are collectively encoded so that the encoding can be performed while reducing the amount of data as compared with the case of independently encoding each dimension of the quantization index vector. It can be effective.

[0148] This is not limited to the case where the quantization index is expressed in the state of three values. For example, when the quantization index is expressed in the state of five values, the three-dimensional sum of 125 is 125 by encoding three dimensions together so that the encoding can be performed in 7 bits = 128 state (within 128 states). . Independently encoding three dimensions requires three bits (8 states) x 3 dimensions = 9 bits. For this reason, two bits can be deleted by encoding three dimensions together.

[0149] Further, when combining the encoded quantization index vectors output from the encoding means 7 (when comparing quantization index vectors extracted from one picture with quantization index vectors extracted from another picture and determining whether or not those pictures are the same). ) Decodes the value of the quantization index for each dimension (for example, in the above example, decodes the quantization index value of +1, 0, -1 for each dimension) and decodes the decoded quantization index. You may calculate an identity measure (the number of dimensions (similarity) to which a quantization index matches, or the number of dimensions (humming distance) which a quantization index does not match) based on a basis.

In addition, the lookup table may be used to perform a combination without decoding to the value of the quantization index for each dimension while being encoded. That is, for each encoded unit, the identity measure (similarity or distance) is stored in the form of a table (lookup table) in advance, and the identity measure (similarity or distance) for each coded unit is obtained by referring to the lookup table. By summating them (for example, by calculating the totals), one can measure the identity of all dimensions.

For example, in the case where the above 5-dimensional units are encoded in one byte (8 bits), the lookup table having a size of 243 × 243 is generated in advance because each 5-dimensional unit is one of 243 states. We can cope by making it. That is, the equality measure between all possible combination states (243 states x 243 states) of the codes of the two five-dimensional units to be compared, that is, the number (similarity) in which the quantization indices in five dimensions coincide, or the quantization index in five dimensions, A non-matching number (humming distance) is calculated in advance and stored as a lookup table of 243 x 243 size. With this table, you can refer to the lookup table for every five-dimensional unit (without decoding to the quantization index for each dimension) to obtain a measure of equality for every five-dimensional unit. For example, in the case where the total number of quantization index vectors is 300 dimensions, since the total of the dimension of the quantization index vector is 300, one byte is encoded by 60 bytes in total, the lookup table is referred to 60 times, and the identity measure of each five-dimensional unit is obtained. The totality (300 dimensions) of the identity measure (similarity or humming distance) can be calculated by adding them together. By using a lookup table, the combination (calculation of the identity measure) can be performed without decoding by the quantization index for each dimension, thereby reducing the processing cost at the combination (calculation of the identity measure), There is an effect that the combination of (calculation of identity measure) becomes possible.

In addition, the identity measure between the two quantization index vectors is not simply calculated as a number of dimensions (similarity) in which the quantization indexes match or a number of dimensions (humming distance) in which the quantization indexes are inconsistent. Even when calculating based on a complicated calculation formula, a combination (calculation of the sameness measure) can be performed using a lookup table without decoding the quantization index for each dimension. For example, a method of calculating the following identity measure is considered as the identity measure of the quantization index vector whose quantization index is calculated based on the equation (2). First, quantization index vectors of two images are compared with corresponding dimensions, and the number of dimensions where both quantization indexes are not 0 is calculated, and this value is set to A. FIG. Next, in the dimension where both quantization indices are not 0, the number of dimensions to which the quantization indices match is calculated as B (or in the dimension where both quantization indices are not 0), the quantization index Calculates the number of dimensions for which C is non-matching as C). The identity measure is calculated as B / A (or the identity measure is calculated as C / A). However, in the case of A = 0 (that is, when both quantization indices become 0 in all dimensions), the identity measure is set to a predetermined numerical value (for example, 0.5). When employing such a method of calculating the identity measure, it is necessary to calculate two values, a value of A and a value of B (or a value of C). In this case, a lookup table of size 243 × 243 for referring to the value of A for every five dimensions and a lookup table of size 243 × 243 for referring to the value of B (or the value of C) for every five dimensions You can cope by creating it. That is, the value of A (the number of dimensions in which both quantization indices are not zero) and all possible pairing states (243 states × 243 states) of the two five-dimensional units of the sign to be compared. The value of B (or the value of C) between 243 states and 243 states is calculated beforehand. Each is stored as a lookup table having a size of 243 × 243. If so, you can look up the lookup table for every five-dimensional unit (without decoding the quantization index for each dimension) to obtain the values of A and B (or C) for each five-dimensional unit. For example, if the total number of quantization index vectors is 300, the size of A in each 5-dimensional unit is referred to 60 times by referring to the lookup table 60 times because the total number of dimensions of the quantization index vector is 300. The value of A and B in all dimensions (300 dimensions) are obtained by obtaining the values of A and B (or C), and then summing the values of A and B (or C) in 5-dimensional units. We can calculate the value of (or the value of C). And finally, by calculating B / A (or C / A), the identity measure can be calculated. In this way, the identity measure is not simply calculated as the number of dimensions (similarity) to which the quantization indexes match, or as the number of dimensions (humming distance) where the quantization indexes are inconsistent, but rather based on a more complex calculation. Even in the case, with reference to the lookup tables, a combination (calculation of the identity measure) is possible without decoding the value for each dimension or for the quantization index. This can reduce the processing cost at the time of the combination (calculation of the identity measure), and has the effect of enabling a high speed combination (calculation of the identity measure).

[Effect of Fifth Embodiment]

It is possible to output the quantization index vector as a smaller data amount.

[0154] Next, the sixth to eighth embodiments of the present invention will be described.

[Sixth Embodiment]

In the sixth embodiment, the number of dimensions of the feature vector to be extracted is 300 dimensions (300th dimension from the first dimension).

[0156] In the sixth embodiment, the extraction regions (first extraction region and second extraction region) for each dimension are constituted by rectangles of various shapes. In the sixth embodiment, the dimensional extraction area information supplied as an input to the extraction area acquisition means 2 is shown in FIG. Fig. 14 shows XY of four corners of a rectangle of extraction areas (first extraction area and second extraction area) for each dimension for an image size of 320 pixels in width x 240 pixels in width, which is a defined image size. Represents a coordinate value. For example, the extraction region of the first dimension may include a first square consisting of four corners of coordinate values (262.000, 163.000), coordinate values (178.068, 230.967), coordinate values (184.594, 67.411), and coordinate values (100.662, 135.378). It is composed of a first extraction area consisting of a quadrangle of the extraction area of the coordinates (161.000, 133.000), coordinate values (156.027, 132.477), coordinate values (164.240, 102.170), coordinate values (159.268, 101.647).

[0157] The extraction area (first extraction area and second extraction area) for each dimension is surrounded by the coordinate values of these four corners with respect to an image normalized to an image size of 320 pixels in width x 240 pixels in width. It is a set of pixels of coordinate values of integer values contained in the area. However, the negative coordinate values contained in the area surrounded by the four corner coordinate values are not included in the extraction area.

In the sixth embodiment, the dimension-specific region feature variable calculation method information supplied as an input to the region feature variable calculating method acquisition unit 5 is shown in FIG. In 6th Embodiment, the average value of the luminance value of the pixel group contained in each extraction area | region (1st extraction area | region and 2nd extraction area | region) is an area feature amount of each extraction area | region in all the dimensions.

In the sixth embodiment, dimensional comparison and quantization method information supplied as an input to the comparison method acquisition means 6 is shown in FIG. In the sixth embodiment, the comparison and quantization method B or the comparison and quantization method G is used for each dimension, and the value of the parameter differs for each dimension. For example, in the first dimension, the comparison and quantization method G is used and the threshold th = D (floor (300 × 5.0 / 100)). Further, for example, in the second dimension, the comparison and quantization method G is used, and the threshold value th = D (floor (300 × 10.0 / 100)). In addition, in the ninth dimension, for example, the comparison and quantization method B is used, and the threshold value th = 3.0.

 [Seventh Embodiment]

In the seventh embodiment, as in the sixth embodiment, the number of dimensions of the feature vector to be extracted is 300 dimensions (300th dimension from the first dimension). In the seventh embodiment, as shown in the sixth embodiment, the information shown in FIG. 14 is used as the dimensional extraction area information supplied as an input to the extraction area acquisition means 2. In the seventh embodiment, the information shown in FIG. 17 is used as the dimension comparison and quantization method information supplied as an input to the comparison method acquisition means 6 as in the sixth embodiment.

[0161] In the seventh embodiment, the dimension-specific region feature amount calculation method information supplied as an input to the region feature variable calculation method acquisition means 5 is shown in FIG. In the seventh embodiment, for each dimension, the average value of the luminance values of the pixel groups included in the extraction region (the first extraction region and the second extraction region), or the percentage luminance value characteristic amount is used, and the same percentage luminance value characteristic is used. Even when an amount is used, the feature amount is different for each dimension. For example, the first dimension is an average value of luminance values of pixels included in the extraction region. For example, the fourth dimension is the percentage luminance characteristic value and Y (floor (N × 20.0 / 100). The eighth dimension is the percentage luminance characteristic value and Y (floor (N × 80.0). / 100).

[Eighth Embodiment]

In the eighth embodiment, the number of dimensions of the feature vector to be extracted is 325 dimensions (the first to 325 dimensions).

In the eighth embodiment, the number of dimensions of the feature vector to be extracted is 325 dimensions (the first to 325 dimensions). In the case of the seventh embodiment, each area is composed of a combination of blocks among 1024 blocks formed by dividing an image into 32 vertical directions and 32 horizontal directions. Here, for each block, as shown in FIG. 28, an index starting from 0 is given from the upper left, and an area is described using this index. Specifically, the rectangular area is expressed as "a-b" using the index "a" of the upper left block and the index "b" of the lower right block. For example, a rectangle composed of four blocks of indexes 0, 1, 32, and 33 is described as 0-33. In the case where the rectangles formed in this way are connected using the symbol "|", the area formed by connecting the rectangles before and after the symbol is represented. For example, 0-33 | 2-67 is an area formed by connecting a rectangle defined by 0-33 and a rectangle defined by 2-67, that is, block numbers 0, 1, 2, 3, 32, The area formed by 33, 34, 35, 66, and 67 is shown.

[0163] FIG. 26 shows regions corresponding to each dimension of the eighth embodiment described in this manner. In FIG. 29, dimensions 325 are described by dividing the regions into types of FIGS. 29A, 29B, 29C, 29D, 29E, 29F, and 29G. Here, the type of region means a group consisting of dimensions having a similar region pattern determined by the relative position between the first and second extraction regions or the knitting of the shape.

Specifically, in the case of FIG. 29A, as shown in FIG. 31A, two regions formed by dividing a square consisting of four vertical and horizontal blocks into two equally in the vertical direction and the horizontal direction are formed. It corresponds to the case where it is used as an extraction area. For this reason, the shape of both the 1st, 2nd extraction area | region is a rectangle which consists of 4 vertical blocks and 2 horizontal blocks, or a rectangle which consists of 2 vertical blocks and 4 horizontal blocks. Moreover, looking at the relative position between 1st, 2nd extraction area | region, it exists in the position adjacent to each other so that the vertical side of a rectangle may overlap each other.

In the case of FIG. 29B, as shown in FIG. 31B, two regions formed by combining the upper left and lower right, the upper right and the lower left, respectively, among four squares formed by vertically and horizontally equilateral squares consisting of eight vertical and horizontal blocks. Corresponds to the case where is used as the first and second extraction regions, respectively. For this reason, the shape of both the 1st, 2nd extraction area | region becomes a shape which arrange | positioned two squares which consist of two vertical and horizontal 2 blocks on the diagonal of 45 degree | time or 135 degree so that one vertex may share. In addition, in view of the relative position between the regions, the second region exists at the position immediately adjacent to the left side and the bottom of the square on the upper left side of the first region, which constitutes the second region.

In the case of FIG. 29C, as shown in FIG. 31C, the shapes of both the first and second extraction regions are squares consisting of 10 vertical and horizontal blocks. Moreover, when looking at the relative position between the 1st, 2nd extraction area | regions, it exists in the position which is a distance of 10 blocks in both vertical and horizontal directions.

In the case of FIG. 29D, as shown in FIG. 31D, the shapes of both the first and second extraction regions are squares composed of six vertical and horizontal blocks. In addition, looking at the relative position between the first and second extraction regions, the vertical and horizontal regions exist at positions far apart by an integer multiple of 6 blocks.

In the case of FIG. 29E, as shown in FIG. 31E, in the case where two regions formed by dividing a square region into a square of a central portion and two outside thereof are used as the first and second extraction regions. It is considerable. For this reason, the shape of a 2nd extraction area | region is a square of a center part, and the shape of a 1st extraction area | region is a shape which cut out the 2nd extraction area | region from the whole square. Moreover, looking at the relative positional relationship between the regions, the second extraction region exists at the position of the hole in the center of the first extraction region.

In the case of FIG. 29F, as shown in FIG. 31F, the shape of the first extraction region is 6 blocks long, a rectangle defined by 10 blocks horizontally, and the second extraction region is 10 blocks vertically, horizontally. It is a rectangle defined by 6 blocks. Moreover, looking at the relative positional relationship between the 1st, 2nd extraction area | region, it arrange | positions so that a center position may correspond.

In the case of FIG. 29G, as shown in FIG. 31G, three vertical sides are equally formed by forming a rectangle consisting of four vertical blocks and 12 horizontal blocks or a rectangle consisting of 12 vertical blocks and 4 horizontal blocks. This corresponds to a case where the central square and two other regions are used as the first and second extraction regions. For this reason, the shape of a 1st extraction area | region is a shape which arrange | positioned two squares which consist of 4 vertical and horizontal blocks, and 4 blocks apart vertically and horizontally, and the shape of a 2nd extraction area | region is a square consisting of 4 vertical and horizontal blocks. In addition, in view of the relative positions of the regions, a second extraction region exists between the first extraction regions.

Hereinafter, the region types of FIGS. 29A, 29B, 29C, 29D, 29E, 29F, and 29G are referred to as region type a, region type b, region type c, region type d, and region type, respectively. Let e be called area type f and area type g.

In the eighth embodiment, the average of the luminance values is calculated as the area feature amounts in each area shown in FIG. 29, and the feature amounts of each dimension are calculated. Of course, instead of the average value of the luminance values, it is possible to obtain values extracted by the above-described various extraction methods such as median values and maximum values as the area feature amounts.

In quantization of the feature quantities in each dimension, quantization is performed by setting a threshold value for each type of region described above. For example, in the case of quantizing the feature amount to three values according to equation 2, the threshold value th of the quantization is determined so that the occurrence ratio of 0, 1,-1 is equalized for each type of region, and quantization is performed. do. Specifically, the method described in paragraph 0085 is applied for each region type, where P = 33.333% and N represents the number of dimensions for each region type, thereby obtaining a threshold value th. For example, in the case of the area type a, since N = 113, the threshold value is calculated by th = D (floor (113 × 33.333 / 100)) = D (37). Note that D (i) (i = 0, 1, ..., N-1) is a set sorted in ascending order of the absolute value of the difference value in the 113th dimension from the first dimension corresponding to the area type a. In this case, the index corresponding to the threshold becomes 37. Similarly, as shown in Fig. 30, the index corresponding to the threshold value can be obtained also for other area types. By obtaining the threshold value for each region type in this manner, the probability of occurrence of 0, 1, -1 in each dimension can be made uniform as compared with the case where the threshold value is determined as a whole, and the identification capability is improved. Of course, it is possible to perform quantization by the various other quantization methods described above.

In the case of the eighth embodiment, it is also possible to first calculate the representative value (for example, the average value of the luminance values of the pixels in the block) for each block shown in FIG. 28, and then extract the region feature amount. Do. This makes it possible to extract at higher speed than in the case of extracting the region feature amount directly from all the pixels in the region. Moreover, the extraction region of each region type has symmetry as a whole. For this reason, even when the right and left sides of the image are inverted or the top and bottom are reversed, the characteristic amount of the original image can be restored from the feature amounts extracted from the left and right or upside down images by appropriately changing the correspondence and the sign of the dimension. Can be. For this reason, the combination may be performed on the image in which the left and the right or the top and bottom are reversed.

[Embodiment of Combination Means]

Next, a combination means for performing a combination between the quantization index vectors output in the present invention will be described using a block diagram.

Referring to FIG. 21, there is shown a block diagram of a combining means for performing a combination between the quantization index vectors output in the present invention, wherein the combining means 100 is a dimension determining means 101, a quantization value obtaining means. 102 and 103, and scale calculation means 104.

[0177] The dimension determining means 101 is connected to the quantized value obtaining means 102 and 103 and outputs the determined dimension information. The quantization value acquisition means 102 obtains the quantization index value of the dimension input from the dimension determination means 101 from the first quantization index vector, and outputs it to the scale calculation means 104 as the first quantization index value. . The quantization value acquisition means 103 obtains the quantization index value of the dimension input from the dimension determination means 101 from the second quantization index vector, and outputs it to the scale calculation means 104 as the second quantization index value. The scale calculation means 104 calculates a scale indicating identity from the first and second quantization index values output from the quantization value acquisition means 102 and 103, respectively, and outputs it.

[0178] Next, the operation of the combining means 100 of FIG. 21 will be described.

[0179] First, the combining means 100 receives a first quantization index vector, which is a quantization index vector extracted from a first image, and a second quantization index vector, which is a quantization index vector extracted from a second image. . The input first and second quantization index vectors are input to the quantization value acquisition means 102 and 103, respectively.

[0180] Dimensional information output from the dimension determination means 101 is also input to the quantization value acquisition means 102 and 103. The dimension determining means 101 sequentially outputs information specifying each dimension of the quantization index vector, which is an N-dimensional vector. The order of output does not necessarily need to be increased from 1 to N one by one, and may be any order as long as the order from 1 to N is specified without being insufficient.

In the quantization value acquisition means 102 and 103, the quantization index value of the dimension specified in the dimensional information output from the dimension determination means 101 is obtained from the input quantization index vector, and the acquired quantization index value is measured. Output to 104.

[0182] The scale calculating means 104 compares the first quantization index value and the second quantization index value output from the quantization value obtaining means 102. This comparison is performed for each dimension to calculate a similar measure (or distance measure) between the first and second quantization index vectors as the identity measure.

The obtained identity measure is compared with a predetermined threshold value to determine the identity. If the identity measure is a measure representing the similarity value, it is determined that the measure value is the same when the measure value is equal to or greater than the threshold value. On the other hand, when the identity measure is a measure indicating distance, it is determined that the measure value is the same when the measure value is equal to or less than the threshold value.

Next, the operation of the combining means 100 of FIG. 21 will be described using a flowchart. First, the operation in the case of using the similarity value as the identity measure will be described.

FIG. 22 is a flowchart showing the operation of the combining means 100. As shown in FIG. In the flowchart of FIG. 22, the dimension (number) of a quantization index vector is represented by "n", and there exists a total N dimension from 1 to N. FIG. Moreover, the variable which calculates a similarity value is shown by B. FIG.

[0186] First, the dimension determination means 101 determines dimension 1 as the first dimension of the combined quantization index vector (n = 1), supplies it to the quantization value acquisition means 102 and 103, and calculates the scale. In the means 104, the variable B is set to 0 (step S100).

[0187] Next, in the quantization value obtaining means 102 and 103, the first quantization index value and the second quantization index value of the dimension n are obtained from the first quantization index vector and the second quantization index vector. The solution is supplied to the scale calculation means 104 (step S102).

[0188] Next, the scale calculation means 104 calculates a similarity value ΔB between the first and second quantization index values and the feature amount corresponding to each quantization index (step S104). . For example, when the quantization indices coincide, ΔB = 1, otherwise ΔB = 0. Alternatively, the representative value of the feature amount before quantization may be calculated from the quantization index, and a value that increases as the difference between the representative values becomes smaller may be used as ΔB. In this case, instead of calculating the representative value of the feature quantities and obtaining the difference, it is necessary to maintain a table in which the value of ΔB is obtained by the combination of the quantization index values, and to directly obtain the value of ΔB using this table from the combination of the quantization index values. It is possible.

[0189] Next, the value of ΔB is added to the variable B (step S106). At this time, when the value of ΔB is 0, instead of adding 0 to the variable B, it is possible to control not to add.

[0190] Next, it is checked whether the number n of the dimension has reached the dimension number N (step S108), and if it does not reach, the process proceeds to step S112, and when it reaches, the variable B at that time. Is output as an identity measure (a measure representing the similarity value) (step S110), and the processing ends.

In step 112, the dimension determining means 101 determines the next dimension from n = n + 1 as the dimension of the quantization index to be obtained and supplies it to the quantization value obtaining means 102 and 103. The process then proceeds to step S102 again.

In addition, although extraction processing is performed in order from dimension 1 to N here, it is noted that an arbitrary order may be taken without being limited to this.

[0193] Next, operation in the case of using distance as the identity measure will be described.

23 is another flowchart showing the operation of the combining means 100. As shown in FIG. Also in the flowchart of FIG. 23, the dimension (number) of a quantization index vector is represented by n, and there exists a total N dimension from 1 to N. FIG. In addition, the variable which calculates a distance measure is shown by C. FIG.

The basic flow is the same as in the case of FIG. 22, but FIG. 23 differs in that the steps S100, S104, S106, and S110 are replaced by the steps S200, S204, S206, and S210, respectively.

First, in step S200, the dimension determination means 101 determines dimension 1 as the first dimension of the combined quantization index vector (n = 1) and supplies it to the quantization value acquisition means 102 and 103; In addition, in the scale calculation means 104, the variable C is set to zero.

In step S204, the scale calculation means 104 calculates the distance ΔC between the feature amounts corresponding to the respective quantization indexes from the first quantization index value and the second quantization index value. For example, when the quantization indexes coincide, ΔC = 0, otherwise ΔC = 1. The representative value of the feature amount before quantization may be calculated from the quantization index, and a smaller value as the difference between the representative values becomes smaller may be used as ΔC. In this case, instead of calculating the representative value of the feature quantities to obtain the difference, it is also possible to maintain a table from which the value of ΔC is obtained from the combination of the quantization index values, and to directly obtain the value of ΔC from this combination of the quantization index values. Do.

In step S206, the value of ΔC is added to the variable C. At this time, when the value of ΔC is 0, instead of adding 0 to the variable C, it is possible to control not to add.

In step S210, the value of the variable C at that time is output as an identity measure (measurement indicating distance), and the process ends.

Other steps are the same as in the case of FIG. However, if the number n of the dimension reaches the dimension number N in step S108, the process proceeds to step S210.

In addition, although extraction processing is performed in order from dimension 1 to N here, it is noted that an arbitrary order may be taken without being limited to this.

[0202] Next, an operation in the case where the similarity value is used as the identity measure except for the dimension in which both quantization index values are 0 as the first quantization index value and the second quantization index value will be described. .

24 is another flowchart illustrating the operation of the combining means 100. Also in the flowchart of FIG. 24, the dimension (number) of a quantization index vector is represented by "n", and there exists a total N dimension from 1 to N. FIG. Moreover, the variable which calculates similarity value is shown by B, and the variable for counting the dimension whose "quantization index of both is not 0" is represented by A.

[0204] First, the dimension determination means 101 determines dimension 1 as the first dimension of the combined quantization index vector (n = 1), supplies it to the quantization value acquisition means 102 and 103, and calculates the scale. In the means 104, the variables A and B are set to 0 (step S300), and the process proceeds to the next step S102.

Step S102 is the same as the case of FIG. 22, and after the step S102 ends, the process proceeds to step S314.

In step S314, the scale calculation means 104 checks whether both the first quantization index value and the second quantization index value are zero. If both values are zero, the process proceeds to step S108. If either value is not zero, the value of variable A is incremented by one (step S316), and the process proceeds to step S104.

[0207] The processing of steps S104, S106, S108, and S112 is the same as that of FIG. If the number n of dimensions reaches the dimension number N in step S108, the process proceeds to step S310.

In step S310, the scale calculation means 104 calculates a value of B / A, outputs it as an identity measure, and ends the processing. However, in the case of A = 0, the scale calculation means 104 outputs a predetermined value (for example, 0.5).

In addition, although extraction processing is performed in order from dimension 1 to N here, it is noted that an arbitrary order may be taken without being limited to this.

[Other Embodiments of Combination Means]

Next, another embodiment of the combining means for performing the combination between the quantization index vectors output in the present invention will be described using a block diagram.

Referring to FIG. 25, there is shown a block diagram of a combining means 200 for performing a combination between quantization index vectors output in the present invention, the combining means 200 being a code determining means 201, a sign. Value obtaining means (202 and 203) and scale calculating means (204).

[0212] The code determining means 201 is connected to the code value obtaining means 202 and 203 and outputs the determined code designation information. The code value obtaining unit 202 obtains the value of the code determined by the code designation information input from the code determining unit 201 from the first coded quantization index vector, and measures the scale as the first code value. Output to 204. The code value obtaining unit 203 obtains the value of the code determined by the code designation information input from the code determining unit 201 from the second coded quantization index vector, and measures the scale as the second code value. Output to 204. The scale calculation means 204 calculates a scale indicating the identity from the first and second code values output from the code value acquisition means 202 and 203, respectively, and outputs it.

[0213] Next, the operation of the combining means 200 of FIG. 25 will be described.

[0214] First, the combining means 200 encodes a first coded quantization index vector, which is a vector generated by encoding a quantization index vector extracted from a first image, and a quantization index vector extracted from a second image. A second coded quantization index vector, which is a generated vector, is input. Here, the coded quantization index vector is a code string consisting of codes obtained by collectively encoding the quantization index values of the quantization index vectors for a plurality of dimensions. As described in paragraph 0142, when the feature quantities of each dimension of the feature vector are quantized to three values and collectively encoded for five dimensions, one code is generated for every five dimensions. Therefore, when the number of dimensions of the feature vector is N, N / 5 codes are generated. In this case, the coded quantization index vector is a code string consisting of N / 5 codes.

[0215] The input first and second coded quantization index vectors are input to the code value obtaining means 202 and 203, respectively.

[0216] Sign designation information output from the sign determining means 201 is also input to the sign value obtaining means 202 and 203. The code determination means 201 sequentially outputs information specifying each code in the code string. If the number of codes in the code string is M (M = N / 5 in the example described above), the output order does not necessarily need to increase from 1 to M one by one, so that the values from 1 to M are not insufficient. As long as the order is specified, any order is possible.

In the code value obtaining means (202 and 203), from the input coded quantization index vector, the value of the code specified in the code designation information output from the code determining means 201 is obtained, and the code value obtained is scaled. Output to the means 204.

In the scale calculation means 204, the first code value output from the code acquisition means 201 and 202 is compared with the second code value. At this time, the code values are directly compared without decoding to the quantization index value. As described in paragraph 0150 to paragraph 0152, a lookup table capable of obtaining an identity measure between their codes from two code values is prepared, and the identity measure is calculated in units of codes using this. This process is performed for each code, and the similarity measure (or distance measure) between the first and second code values is calculated as the identity measure.

[0219] The measured identity measure is compared with a predetermined threshold value to determine the identity. If the identity measure is a measure representing the similarity value, it is determined that the measure value is the same when the measure value is equal to or greater than the threshold value. On the other hand, when the identity measure is a measure indicating distance, it is determined that the measure value is the same when the measure value is equal to or less than the threshold value.

[0220] Next, the operation of the combining means 200 of FIG. 25 will be described using a flow chart. First, the operation in the case of using the similarity value as the identity measure will be described.

FIG. 26 is a flowchart showing the operation of the combining means 200. As shown in FIG. In the flowchart of FIG. 26, the code | symbol number of the coded quantization index vector is represented by "m", and there exist M codes from 1 to M in total. Moreover, the variable which calculates a similarity value is shown by B. FIG.

[0222] First, the code determination means 201 determines to acquire the first code as the first code of the coded quantization index vector to be combined (m = 1), and gives the code value acquisition means (202, 203). In addition to supplying, the variable B is set to zero by the scale calculation means 204. (Step S600).

[0223] Next, in the code value obtaining means 202 and 203, the m-th first code value and the second code value are obtained from the first coded quantization index vector and the second coded quantization index vector. It acquires and supplies to the scale calculation means 204 (step S602).

[0224] Next, the scale calculation means 204 looks up, from the first code value and the second code value, the similarity value ΔB between the feature quantities of the plurality of dimensions corresponding to the respective code values, as described in paragraph 0150. It calculates by referring to a table (step S604).

[0225] Next, the value of ΔB is added to the variable B (step S106). At this time, when the value of ΔB is 0, instead of adding 0 to the variable B, it is also possible to control not to add.

[0226] Next, it is checked whether the number m of the code has reached the number M of the codes (step S608). If not, the process proceeds to step S612. The value is output as an identity measure (a measure representing the similarity value) (step S110), and the process ends.

In step 612, the code determination means 201 determines the number of the next code from m = m + 1 as a dimension for obtaining the quantization index, and gives the code value acquisition means 202,203 to the code designation information. Supply. The process then proceeds to Step S602 again.

In addition, although extraction processing is performed in order from code | symbol number 1 to M here, it is noted that the order is not limited to this, and arbitrary order may also be taken. In addition, although the case where the similarity value is computed was demonstrated here, similarly, the distance scale can also be computed as an identity measure. In this case, the lookup table is fitted to maintain the distance scale instead of the similarity value.

FIG. 27 is another flowchart showing the operation of the combining means 200. As shown in FIG. Also in the flowchart of FIG. 27, the code | symbol number of the coded quantization index vector is represented by m, and there exist M code | symbols from 1 to M in total. Moreover, the variable which calculates similarity value is shown by B, and the variable for counting the dimension whose "quantization index of both is not 0" is represented by A.

First, the code determination means 201 determines to acquire the first code as the first code of the coded quantization index vector to be combined (m = 1), and to the code value acquisition means 202 and 203. While supplying, the variables A and B are set to 0 in the scale calculating means 204 (step S700), and the process proceeds to step S602.

Step S602 is the same as in the case of FIG. 26, and the process proceeds to step S714 after the end.

In step S714, in the scale calculation means 204, in the dimension of the feature vectors corresponding to the sign value from the first sign value and the second sign value, the value of both is not zero. Examine how many dimensions there are. Set the number of these dimensions to ΔA. This can also be calculated by using a lookup table describing the relationship between the sign value and ΔA, as described in paragraph 0152.

[0233] Next, the value of ΔA is added to the variable A (step S716). At this time, when the value of ΔA is 0, instead of adding 0 to the variable A, it is possible to control not to add.

[0234] The processing of steps S604, S106, S608, and S612 is the same as the case of FIG. If the code number m reaches the number M of codes in step S608, the flow advances to step S310.

In step S310, the scale calculation means 204 calculates a value of B / A, outputs it as an identity measure, and ends the process. However, in the case of A = 0, the scale calculation means 204 outputs a predetermined value (for example, 0.5).

In addition, although extraction processing is performed in order from the code | symbol number m to M here, in order, it is not limited to this, It is arbitrary. In addition, although the case where the similarity value is computed was demonstrated here, similarly, the distance scale can also be computed as an identity measure. In this case, the lookup table is fitted to maintain the distance scale instead of the similarity value.

Although the embodiment of the present invention has been described above, the present invention is not limited to the above-described embodiment. The structure and details of the present invention can be variously modified by those skilled in the art within the scope of the present invention. In addition, the image identifier extracting apparatus of the present invention can be realized not only by its hardware but also by a computer and a program. Such a program is provided in a form recorded in a computer readable recording medium such as a magnetic disk or a semiconductor memory, and is read into the computer at the start of the computer, for example, to control the operation of the computer, thereby implementing the computer as described above. It functions as dimensional determination means, extraction area acquisition means, area feature variable calculation means, comparison means, area feature variable calculation method acquisition means, and comparison method acquisition means in the form.

[0238] Furthermore, the present invention provides the benefit of claiming priority based on Japanese Patent Application No. 2009-01022, filed on March 13, 2009, and Japanese Patent Application No. 2009-097864, filed on April 14, 2009. Since it is a perfume, all the content described in the said patent application shall be included in this specification.

[0239] 1... Dimension
2… Extraction area acquisition means
3, 3 A... Domain feature calculation unit
31, 31 A... First region feature variable calculating means
32, 32 A... Second area feature variable calculating means
4, 4 B... Comparison means
41... Comparison measure
42, 44... Quantization means
43 ... Difference calculation means
45... Quantization Boundary Determination Means
5... Area feature variable calculation method acquisition means
6... Comparison method acquisition means
7 ... Encoding means

Claims (57)

Both of the pairing of the shape of two paired subregions and the relative position between the two paired subregions includes one or more partial region pairs that differ from at least one other subregion pair. Image identifier generating means for extracting an area feature amount from each partial area of an image according to a plurality of partial area pairs, and generating an image identifier for use in identifying the image based on the extracted area feature amounts of each partial area; ; And
And encoding means for encoding the image identifier. The method of claim 1,
And the encoding means encodes the image identifier so that the amount of data is small. The method according to claim 1 or 2,
And the encoding means encodes the image identifier into a uniquely decodable format. The method according to any one of claims 1 to 3,
An image identifier extracting apparatus in which absolute positions differ in both a pair of partial region pairs and a relative position between two paired partial regions in the same partial region pair. The method according to any one of claims 1 to 4,
The image identifier generating means extracts an area feature amount from both of the paired two partial regions, calculates an image identifier element using the extracted area feature amount, and extracts the picture identifier that is a set of the picture identifier elements. An image identifier extraction device to generate. 6. The method according to any one of claims 1 to 5,
The image identifier generating means extracts an area feature amount from both of the paired two partial regions, calculates an image identifier element from the difference value of the extracted area feature amounts, and is the set of the picture identifier elements. An image identifier extraction device for generating an identifier. The method according to claim 6,
And the image identifier generating means quantizes the difference value of both area feature amounts to calculate the image identifier element. The method according to any one of claims 1 to 6,
And the image identifier generating means quantizes the area feature amounts to calculate the image identifier. The method according to claim 7 or 8,
And the image identifier generating means determines the quantization boundary used for the quantization based on the distribution of the difference values between both region feature amounts for each partial region pair. The method according to any one of claims 7 to 9,
And the image identifier generating means determines the quantization boundary used for the quantization so that the ratios of the partial region pairs each classified into different quantization values are almost equal. The method according to any one of claims 7 to 10,
And the image identifier generating means performs quantization with three values. The method of claim 11,
The three values include a first quantized value, a second quantized value that is less than the first quantized value, and a third quantized value that is less than the second quantized value, wherein the first quantized value And a difference between the second quantized value and a difference between the second quantized value and the third quantized value is equal. The method according to claim 11 or 12,
In the image identifier generating means, the region feature amount extracted from the first partial region of the two partial regions constituting the partial region pair is V1, the region feature quantity extracted from the second partial region is V2, and the threshold value of the quantization boundary is th If the value obtained by subtracting V2 from V1 is greater than th, a first quantized value is generated. If the value obtained by subtracting V2 from V1 is less than -th, a third quantized value is generated. And generating a second quantized value if the absolute value of the value obtained by subtracting is equal to or less than th. The method of claim 13,
And the image identifier generating means sorts the absolute value of the difference value for each partial region pair and uses the value at a position of a predetermined ratio from the upper or lower portion as the threshold value th. The method according to any one of claims 11 to 14,
And the encoding means encodes a quantized value of five partial region pairs as one unit. The method according to any one of claims 11 to 15,
And the encoding means encodes the quantized values of the five partial region pairs as one byte of information. The method according to any one of claims 11 to 15,
And said encoding means maps 243 different sets of quantization values of five partial region pairs to values from 0 to 242 expressed in one byte. The method of claim 17,
When performing the mapping, the encoding means treats the smallest quantized value as 0, the next smallest quantized value as 1, and the largest quantized value as 2, among the three values of the quantized value, , X ₄ , X ₃ , X ₂ , X ₁ , X ₀ , image identifier extraction to calculate the encoded value Z as Z = 81 × X ₄ + 27 × X ₃ + 9 × X ₂ + 3 × X ₁ + X ₀ Device. The method according to any one of claims 5 to 7,
And the encoding means encodes a plurality of image identifier elements as one unit. 20. The method according to any one of claims 1 to 19,
And the image identifier generating means extracts a representative value of pixel values of each partial region of the image as an area feature amount of the partial region. The method of claim 1, wherein
And the image identifier generating means extracts an average luminance value of each partial region of the image as an area feature amount of the partial region. A combining apparatus that performs combining using an image identifier generated by the image identifier extracting apparatus according to any one of claims 1 to 21. An identification apparatus for performing identification using an image identifier generated by the image identifier extracting apparatus according to any one of claims 1 to 21. The first coded picture identifier and the second coded picture, using a first coded picture identifier for identifying a first picture and a second coded picture identifier for identifying a second picture. And a combination between the first coded picture identifier and the second coded picture identifier without decoding an identifier. The method of claim 24,
In the combination, an image identifier combination device in which an identity measure indicating a degree to which the first image and the second image are equal is calculated. The method of claim 24 or 25,
For each coded unit, a lookup table in which the distance or similarity value between the code value of the first coded picture identifier and the code value of the second coded picture identifier calculates and records the distance or similarity value between the code values in advance. And a combination is performed. The method of claim 26,
And the distance between the code values is a humming distance between decoding information that decodes the code value. The method of claim 26,
When the first coded picture identifier and the second coded picture identifier are encoded to be encoded in one byte every five elements having three values per one element, the lookup table is a 243 × 243 scheme. And a distance or similarity of the two. Both of the pairing of the shape of two paired subregions and the relative position between the two paired subregions includes one or more partial region pairs that differ from at least one other subregion pair. Extracting an area feature amount from each partial area of the image according to a plurality of partial area pairs, and generating an image identifier for use in identifying the image based on the extracted area feature amounts of each partial area; And
Encoding the picture identifier. The method of claim 29,
In the encoding step, the image identifier extraction method encodes the image identifier so that the amount of data decreases. 32. The method according to claim 29 or 30,
And the encoding step encodes the image identifier in a uniquely decodable format. The method of any one of claims 29 to 31,
2. The method of extracting an image identifier, in which the absolute position differs in a partial region pair in which both a pair of shapes of paired two partial regions and a relative position of two paired partial regions are the same. The method according to any one of claims 29 to 32,
The generating of the image identifier comprises: extracting an area feature amount from both of the paired two partial regions, calculating an image identifier element using the extracted area feature amount, and the set of image identifier elements Generating the picture identifier. The method according to any one of claims 29 to 33,
The generating of the image identifier may include extracting an area feature amount from both of the paired two partial regions, calculating an image identifier element from the difference between the extracted area feature amounts, and the image identifier element. Generating the picture identifier which is a set of pictures. 35. The method of claim 34,
And generating the picture identifier comprises quantizing the difference of both area feature amounts to calculate the picture identifier element. The method according to any one of claims 29 to 34, wherein
And generating the image identifier comprises calculating the image identifier by quantizing the area feature amount. The method of claim 35 or 36,
The generating of the picture identifier includes determining a quantization boundary used for the quantization based on a distribution of difference values between both area feature amounts for each partial region pair. The method according to any one of claims 35 to 37,
The generating of the image identifier includes determining a quantization boundary used for the quantization such that the ratio of the partial region pairs each classified into different quantization values is nearly equal. The method according to any one of claims 35 to 38,
Generating the picture identifier comprises performing quantization with three values. The method of claim 39,
The three values include a first quantized value, a second quantized value that is less than the first quantized value, and a third quantized value that is less than the second quantized value, wherein the first quantized value And a difference between the second quantized value and a difference between the second quantized value and the third quantized value is equal. 41. The method of claim 39 or 40 wherein
The generating of the image identifier includes: a threshold value of a region feature amount extracted from the first partial region of the two partial regions constituting the partial region pair is V1, an area feature quantity extracted from the second partial region of V2, and a quantization boundary; Is th, subtracting V2 from V1 to generate a first quantized value if greater than th, and subtracting V2 from V1 to generate a third quantized value if less than -th. And generating a second quantized value if the absolute value of the value obtained by subtracting V2 from V1 is less than or equal to th. 42. The method of claim 41 wherein
Generating the image identifier includes sorting the absolute value of the difference value for each of the partial region pairs, and using the value at a predetermined ratio position from a higher or lower position as the threshold value th. Identifier extraction method. 43. The compound of any one of claims 39 to 42 wherein
The encoding step includes encoding the quantized values of the five partial region pairs as one unit. The method according to any one of claims 39 to 43,
The encoding step includes encoding a quantized value of five partial region pairs as one byte of information. The method according to any one of claims 39 to 43,
And the encoding step includes mapping 243 different sets of quantization values of five partial region pairs from 0 to 242 expressed in one byte. The method of claim 45,
The encoding may include treating the smallest quantized value as 0, the next smallest quantized value as 1, and the largest quantized value as 2, when performing the mapping, and 5 partial region pairs. If the quantization value of is X ₄ , X ₃ , X ₂ , X ₁ , X ₀ , calculating the coded value Z as Z = 81 × X ₄ + 27 × X ₃ + 9 × X ₂ + 3 × X ₁ + X ₀ . And an image identifier extraction method. The method according to any one of claims 33 to 35,
The encoding step includes encoding a plurality of image identifier elements as one unit. The method according to any one of claims 29 to 47,
The generating of the image identifier includes extracting a representative value of the pixel value of each partial region of the image as an area feature amount of the partial region. 49. The method of any of claims 29 to 48,
The generating of the image identifier includes extracting an average luminance value of each partial region of the image as an area feature amount of the partial region. A combining method, wherein combining is performed using an image identifier generated by the image identifier extracting method according to any one of claims 29 to 49. An identification method which performs identification using an image identifier generated by the image identifier extraction method according to any one of claims 29 to 49. The first coded picture identifier and the second coded picture, using a first coded picture identifier for identifying a first picture and a second coded picture identifier for identifying a second picture. And performing a combination between the first coded picture identifier and the second coded picture identifier without decoding the identifier. The method of claim 52, wherein
In the combination, an image identifier combination method is calculated in which an identity measure indicating a degree to which the first image and the second image are the same. The method of claim 52 or 53,
For each coded unit, the distance or similarity value between the code value of the first coded picture identifier and the code value of the second coded picture identifier calculates a distance or similarity value between the code values in advance. An image identifier combining method, obtained from a recorded lookup table, and combining is performed. The method of claim 54, wherein
And the distance between the code values is a humming distance between decoding information obtained by decoding the code value. The method of claim 54, wherein
When the first coded picture identifier and the second coded picture identifier are encoded to be encoded in one byte every five elements having three values per one element, the lookup table is a 243 × 243 scheme. And recording the distance or the similarity of the. The computer,
The plurality of images in the image, wherein both the pairing of the shape of the paired two subregions and the relative positions of the paired two subregions comprise at least one subregion pair that differs from at least one other subregion pair. Image identifier generating means for extracting an area feature amount from each partial area of the image according to a partial area pair of and generating an image identifier to be used for identification of the image based on the extracted area feature amount of each partial area; And
And a program for functioning as encoding means for encoding the image identifier.

Images